om2010 tpaper all -...
TRANSCRIPT
Ontology Matching
OM-2010
Proceedings of the ISWC Workshop
Introduction
Ontology matching1 is a key interoperability enabler for the Semantic Web, aswell as a useful tactic in some classical data integration tasks. It takes theontologies as input and determines as output an alignment, that is, a set ofcorrespondences between the semantically related entities of those ontologies.These correspondences can be used for various tasks, such as ontology mergingand data translation. Thus, matching ontologies enables the knowledge anddata expressed in the matched ontologies to interoperate.
The workshop has two goals:
• To bring together leaders from academia, industry and user institutionsto assess how academic advances are addressing real-world requirements.The workshop strives to improve academic awareness of industrial and finaluser needs, and therefore, direct research towards those needs. Simulta-neously, the workshop serves to inform industry and user representativesabout existing research efforts that may meet their requirements. Theworkshop also investigates how the ontology matching technology is goingto evolve.
• To conduct an extensive and rigorous evaluation of ontology matching ap-proaches through the OAEI (Ontology Alignment Evaluation Initiative)2010 campaign2. The particular focus of this year’s OAEI campaign is onreal-world specific matching tasks involving, e.g., biomedical ontologiesand open linked data. Thus, the ontology matching evaluation initiativeitself provides a solid ground for discussion of how well the current ap-proaches are meeting business needs.
We received 29 submissions for the technical track of the workshop. Theprogram committee selected 7 submissions for oral presentation and 13 submis-sions for poster presentation. 15 matching systems participated in this year’sOAEI campaign. Further information about the Ontology Matching workshopcan be found at: http://om2010.ontologymatching.org/.
1http://www.ontologymatching.org/2http://oaei.ontologymatching.org/2010
i
Acknowledgments. We thank all members of the program committee, au-thors and local organizers for their efforts. We appreciate support from theTrentino as a Lab (TasLab)3 initiative of the European Network of the LivingLabs4 at Informatica Trentina SpA5, the EU SEALS (Semantic Evaluation atLarge Scale)6 project and the Semantic Valley7 initiative.
Pavel ShvaikoJerome EuzenatFausto GiunchigliaHeiner StuckenschmidtMing MaoIsabel Cruz
November 2010
3http://www.taslab.eu4http://www.openlivinglabs.eu5http://www.infotn.it6http://www.seals-project.eu7http://www.semanticvalley.org/index_eng.htm
ii
Organization
Organizing Committee
Pavel Shvaiko, TasLab, Informatica Trentina SpA, ItalyJerome Euzenat, INRIA & LIG, FranceFausto Giunchiglia, University of Trento, ItalyHeiner Stuckenschmidt, University of Mannheim, GermanyMing Mao, SAP Labs, USAIsabel Cruz, The University of Illinois at Chicago, USA
Program Committee
Paolo Besana, Universite de Rennes 1, FranceOlivier Bodenreider, National Library of Medicine, USAMarco Combetto, Informatica Trentina, ItalyJerome David, Universite Pierre Mendes-France, INRIA & LIG, FranceAnHai Doan, University of Wisconsin and Kosmix Corp., USAAlfio Ferrara, University of Milan, ItalyTom Heath, Talis, UKWei Hu, Nanjing University, ChinaRyutaro Ichise, National Institute of Informatics, JapanAntoine Isaac, Vrije Universiteit Amsterdam & Europeana, NetherlandsKrzysztof Janowicz, Pennsylvania State University, USABin He, IBM, USAYannis Kalfoglou, Ricoh Europe plc, UKMonika Lanzenberger, Vienna University of Technology, AustriaPatrick Lambrix, Linkopings Universitet, SwedenMaurizio Lenzerini, University of Rome La Sapienza, ItalyJuanzi Li, Tsinghua University, ChinaAugusto Mabboni, Business Process Engineering, ItalyVincenzo Maltese, University of Trento, ItalyFiona McNeill, University of Edinburgh, UKChristian Meilicke, University of Mannheim, GermanyLuca Mion, Informatica Trentina, ItalyPeter Mork, The MITRE Corporation, USAFilippo Nardelli, Cogito, ItalyNatasha Noy, Stanford University, USALeo Obrst, The MITRE Corporation, USAYefei Peng, Google, USAErhard Rahm, University of Leipzig, GermanyFrancois Scharffe, INRIA & LIG, FranceLuciano Serafini, Fondazione Bruno Kessler - IRST, ItalyKavitha Srinivas, IBM, USA
iii
Umberto Straccia, ISTI-C.N.R., ItalyAndrei Tamilin, Fondazione Bruno Kessler - IRST, ItalyCassia Trojahn dos Santos, INRIA & LIG, FranceLorenzino Vaccari, European Commission - Joint Research Center, ItalyLudger van Elst, DFKI, GermanyYannis Velegrakis, University of Trento, ItalyShenghui Wang, Vrije Universiteit Amsterdam, NetherlandsBaoshi Yan, Bosch Research, USARui Zhang, Jilin University, ChinaSongmao Zhang, Chinese Academy of Sciences, China
Additional Reviewers
Zhongli Ding, Google, USAEugenio Di Sciascio, Politecnico di Bari, ItalySongyun Duan, IBM, USAJorn Hees, DFKI, GermanyAnika Gross, University of Leipzig, GermanyNico Lavarini, Cogito, ItalySabine Massmann, University of Leipzig, GermanySalvatore Raunich, University of Leipzig, GermanyZhichun Wang, Tsinghua University, China
iv
Table of Contents
PART 1 - Technical Papers
Linguistic analysis for complex ontology matchingDominique Ritze, Johanna Volker,Christian Meilicke and Ondrej Svab-Zamazal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Lost in translation? Empirical analysis of mapping compositionsfor large ontologiesAnna Tordai, Amir Ghazvinian, Jacco van Ossenbruggen,Mark Musen and Natasha Noy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Chinese whispers and connected alignmentsOliver Kutz, Immanuel Normann, Till Mossakowski and Dirk Walther . . . . . 25
Consistency-driven argumentation for alignment agreementCassia Trojahn and Jerome Euzenat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .37
Alignment based measure of the distance betweenpotentially common parts of lightweight ontologiesAmmar Mechouche, Nathalie Abadie and Sebastien Mustiere . . . . . . . . . . . . . . . 49
Mapping the central LOD ontologies to PROTON upper-level ontologyMariana Damova, Atanas Kiryakov, Kiril Simov and Svetoslav Petrov . . . . . 61
Ontology alignment in the cloudJurgen Bock, Alexander Lenk and Carsten Danschel . . . . . . . . . . . . . . . . . . . . . . . 73
v
PART 2 - OAEI Papers
First results of the Ontology Alignment Evaluation Initiative 2010Jerome Euzenat, Alfio Ferrara, Christian Meilicke, Juan Pane,Francois Scharffe, Pavel Shvaiko, Heiner Stuckenschmidt,Ondrej Svab-Zamazal, Vojtech Svatek, and Cassia Trojahn dos Santos . . . . 85
Using AgreementMaker to align ontologies for OAEI 2010:Isabel Cruz, Cosmin Stroe, Michele Caci, Federico Caimi,Matteo Palmonari, Flavio Palandri Antonelli and Ulas C. Keles . . . . . . . . . . 118
ASMOV: results for OAEI 2010Yves R. Jean-Mary, E. Patrick Shironoshita and Mansur R. Kabuka . . . . . 126
BLOOMS on AgreementMaker: results for OAEI 2010Catia Pesquita, Cosmin Stroe, Isabel Cruz and Francisco Couto . . . . . . . . . . 134
CODI: Combinatorial Optimization for Data Integration:results for OAEI 2010Jan Noessner and Mathias Niepert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Eff2Match results for OAEI 2010Watson Wei Khong Chua and Jung-Jae Kim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
ObjectCoref & Falcon-AO: results for OAEI 2010Wei Hu, Jianfeng Chen, Gong Cheng and Yuzhong Qu . . . . . . . . . . . . . . . . . . . 158
An integrated matching system GeRoMeSuite and SMB:results for OAEI 2010Christoph Quix, Avigdor Gal, Tomer Sagi and David Kensche . . . . . . . . . . . . 166
LN2R a knowledge based reference reconciliation system:OAEI 2010 resultsFatiha Saıs, Nobal Niraula, Nathalie Pernelle andMarie-Christine Rousset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
MapPSO results for OAEI 2010Jurgen Bock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .179
Results of NBJLM for OAEI 2010Song Wang, Gang Wang and Xiaoguang Liu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
RiMOM results for OAEI 2010Zhichun Wang, Xiao Zhang, Lei Hou, Yue Zhao, Juanzi Li,Yu Qi and Jie Tang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Alignment results of SOBOM for OAEI 2010Peigang Xu, Yadong Wang, Liang Cheng and Tianyi Zang . . . . . . . . . . . . . . . 202
vi
TaxoMap alignment and refinement modules: results for OAEI 2010Faycal Hamdi, Brigitte Safar, Nobal Niraula and Chantal Reynaud . . . . . . . 211
vii
PART 3 - Posters
Towards a UMLS-based silver standard for matching biomedical ontologiesErnesto Jimenez-Ruiz, Bernardo Cuenca Grau,Ian Horrocks and Rafael Berlanga . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
From French EHR to NCI ontology via UMLSPaolo Besana, Marc Cuggia, Oussama Zekri,Annabel Bourde and Anita Burgun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
From mappings to modules: using mappingsto identify domain-specific modules in large ontologiesAmir Ghazvinian, Natasha Noy and Mark Musen . . . . . . . . . . . . . . . . . . . . . . . . .223
Harnessing the power of folksonomies for formalontology matching on-the-flyTheodosia Togia, Fiona McNeill and Alan Bundy . . . . . . . . . . . . . . . . . . . . . . . . . 225
Concept abduction for semantic matchmakingin distributed and modular ontologiesViet-Hoang Vu and Nhan Le-Thanh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
LingNet: networking linguistic and terminological ontologiesWim Peters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Aggregation of similarity measures in ontology matchingLihua Zhao and Ryutaro Ichise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Using concept and structure similarities for ontology integrationXiulei Liu, Payam Barnaghi, Klaus Moessner and Jianxin Liao . . . . . . . . . . 233
Flexible bootstrapping-based ontology alignmentPrateek Jain, Pascal Hitzler and Amit Sheth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .235
Semantic matching of ontologiesChristoph Quix, Marko Pascan, Pratanu Roy and David Kensche . . . . . . . . . 237
Ontology mapping neural network: an approachto learning and inferring correspondences among ontologiesYefei Peng, Paul Munro and Ming Mao . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
Towards tailored domain ontologiesCheikh Niang, Beatrice Bouchou and Moussa Lo . . . . . . . . . . . . . . . . . . . . . . . . . 241
Crowd sourcing through social gaming for community drivenontology engineering, results and observationsAlloy Martin Chua, Roland Christian Chua, Arthur Vincent Dychiching,Tinmon Ang, Jose Lloyd Espiritu, Nathalie Rose Lim andDanny Cheng . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
viii
ix
x
Linguistic Analysisfor Complex Ontology Matching
Dominique Ritze1, Johanna Volker1, Christian Meilicke1, and Ondrej Svab-Zamazal2
1University of Mannheim, 2University of Economics, Prague
[email protected],{johanna,christian}@informatik.uni-mannheim.de,
Abstract. Current ontology matching techniques focus on detecting correspon-
dences between atomic concepts and properties. Nevertheless, it is necessary and
possible to detect correspondences between complex concept or property descrip-
tions. In this paper, we demonstrate how complex matching can benefit from nat-
ural language processing techniques, and propose an enriched set of correspon-
dence patterns leveraging linguistic matching conditions. After elaborating on the
integration of methods for the linguistic analysis of textual labels with an existing
framework for detecting complex correspondences, we present the results of an
experimental evaluation on an OAEI dataset. The results of our experiments in-
dicate a large increase of precision as compared to the original approach, which
was based on similarity measures and thresholds.
1 Introduction
Ontology matching can be considered one of the key technologies for efficient knowl-
edge exchange and the successful realization of the Semantic Web. Bridging the gap
between different terminological representations is an indispensable requirement for a
large variety of tools and technologies including, for example, distributed reasoning,
instance migration, and query rewriting in distributed environments.
In the past, ontology matching was commonly considered the task of detecting sim-
ilar or equivalent concepts and properties in two ontologies. However, this view on
ontology matching seems too narrow for many application scenarios, and new require-
ments have motivated several extensions to the original task definition. Among the chal-
lenges of the last Ontology Alignment Evaluation Initiative (OAEI) [4], for example,
we find the task of instance matching as well as a track that aims at the generation of
correspondences expressing subsumption (instead of equivalence) between concepts.
In our work, we suggest to extend the classical way of ontology matching in a
different direction – the generation of correspondences between complex concept and
property descriptions. We refer to these correspondences as complex correspondencesand call the process of generating them as complex ontology matching. Our work is mo-
tivated by the insight that equivalence or even subsumption correspondences between
atomic entities are often not applicable or not expressive enough to capture important
dependencies between ontological entities.
1
This paper is based on our previous approach [11] which we found to have several
disadvantages, including problems related to the precision of the patterns as well as
the prerequisite of a reference alignment as additional input. In order to address these
issues, we modified the original approach in the following way:
– A partial reference alignment is no longer required as input. We generate the align-
ment in a preprocessing step and use it as an anchor alignment later on.
– The complete logic required to express the conditions for generating correspon-
dences is now described declaratively by means of XML. The XML-based specifi-
cation of the matching conditions is interpreted and executed by our tool, while the
concrete implementation remains transparent to the user.
– We changed the output format of our system so that it adheres to the Expres-sive Declarative Ontology Alignment Language (EDOAL) for complex correspon-
dences, that is supported by the alignment API [3].
– We extended our approach by various methods for the linguistic analysis of con-
cept or property labels. According to our experiments, the appropriate use of these
methods results in a significantly increased precision.
The last point refers to the most essential contribution of this paper. In our previous
approach, many matching conditions included similarity thresholds. Thanks to linguis-
tic methods, we can now avoid the need for finding appropriate thresholds. In the re-
mainder of this paper, we show that the use of these methods significantly improves the
quality of complex matching. In particular, we find that the linguistic analysis enables
us to achieve a significantly higher precision with respect to the previously detected [11]
pattern instantiations. Moreover, we present a new correspondence pattern leveraging
linguistic matching conditions and illustrate the advantages of the new approach by
means of concrete examples.
Our paper is structured as follows. In Section 2 we describe our approach to com-
plex matching. We introduce the terminology we adopt, and sketch the core elements of
our algorithm. Section 3 discusses the related work, whereas in Section 4, we describe
the linguistic methods that we apply to detect non-trivial semantic relations between
concepts and properties. The pattern-specific matching conditions, which constitute the
heart of our approach, are presented in Section 5. In Section 6, we report on our evalu-
ation experiments, before concluding in Section 7.
2 Approach
Now we introduce the basics of our approach and explain the terminology we use in the
subsequent sections. First of all, we adopt and slightly simplify the generic terminology
defined in [6]. Thus, we understand an alignment between two ontologies O1 and O2
as a set of correspondences. A correspondence is a triple 〈X,Y, Z〉 where X is an
entity from O1, Y is an entity from O2 and R is a relation such as equivalence or
subsumption. Whenever it is required to refer to the origin of a specific concept, we
write C#i to indicate that C belongs to Oi.
State-of-the-art ontology matching techniques are bound to detect correspondences
as 〈Paper,Article,≡〉 or 〈writes,writesPaper,⊆〉. In the following we describe an ap-
2
proach that allows to detect correspondences where X and Y are complex concept or
property descriptions.
Remember that the power of description logic originates from the ability to build
complex concept and property descriptions from atomic ones, i.e. from concept and
property names. The following listing shows some of the different ways to construct
complex descriptions in description logics (here at the example of SHOIN ).
¬C (atomic negation) ∀P.C (value restriction)
B � C (conjunction) ∃≤nP (at least restriction)
B C (disjunction) ∃≥nP (at most restriction)
{o1, . . . , on} (one of) P−1 (inverse property)
∃P.C (exists restriction)
In this listing, C refers to an arbitrary concept description and P refers to a property
name. We define a correspondence 〈X,Y, Z〉, where X or Y is built according to one
of the rules, as complex correspondence. An alignment that contains a complex corre-spondence as defined as a complex alignment. Note also that several of these rules can
be applied sequentially according to their use in standard description logics.
In the following we will talk about matching conditions and correspondencepatterns. A correspondence pattern describes a special type of complex corre-
spondence. Suppose, for example, that our algorithms detects a correspondence
〈∃earlyRegistered.{true},EarlyRegisteredParticipant,≡〉. This correspondence is a
concrete instantiation of the general correspondence pattern 〈∃P.{true}, C,≡〉, where
P and C denote variables. A correspondence pattern can coincide with one of the con-
struction rules listed above, but can also be compounded of several rules and might
contain constants, as shown in the example.
Fig. 1. Applying matching conditions to detect instances of correspondence patterns in order to
generate a complex alignment.
In our approach we define for each correspondence pattern a set of matching condi-
tions. If these conditions are fulfilled, we generate a concrete instantiation of the pattern
and output a complex correspondence. The corresponding approach is depicted in Fig-
ure 1. First of all we generate by state of the art methods a non-complex alignment.
This alignment is used as a kind of anchor that allows to check if certain relations hold
3
between entities of O1 and O2 such as ”is concept C in O1 a subconcept of concept
D in O2”. The matching conditions used to detect a certain pattern comprise structural
conditions as well as the linguistic conditions are presented in Section 4.
The following section reviews related work in the field of complex ontology match-
ing as well as recent approaches to leveraging linguistic tools and resources in ontology
matching.
3 Related Work
Considering the state-of-the-art in complex ontology matching, we find recent ap-
proaches to be distinguished by three key dimensions: the design, the representation and
discovery of complex correspondences. In [12], complex correspondences are mainly
considered in terms of design and representational aspects. The author proposes align-ment patterns1 as a solution for recurring mismatches raised during the alignment of
two ontologies.2 According to Scharffe [12], complex matching is a task that has to
be performed by a human user (e.g., a domain expert), who can be supported by tem-
plates for capturing complex correspondences. However, similar patterns can also be
exploited by automated matching approaches, as demonstrated in this paper. The align-
ment patterns from [12] are expressed in terms of EDOAL3 [5], an extension of the
alignment format proposed by [3]. It covers concept and property descriptions, concept
restrictions, property value transformations, comparators for restriction over entities,
and variables for representing ontology entities in patterns. In this paper, we adhere to
this expressive language for capturing our correspondence patterns.
Svab-Zamazal et al. [14] consider complex ontology matching as a use case for on-
tology transformation. An ontology is transformed in such a way that it can be more
easily matched with other ontologies. Each transformation is performed by means of a
transformation pattern containing several source and target ontology patterns as well as
an appropriate pattern transformation, which captures the relationships between them.
Each source ontology pattern specifies detection conditions such as structural and nam-
ing conditions. The authors argue that successful non-complex matching applied to the
transformed ontology can be used for finding complex correspondences by tracking
changes back to the original ontology. This approach, however, lacks experimentation.
Further work related to the discovery of complex correspondences relies on ma-
chine learning techniques such as Inductive Logic Programming, for example [9]. This
type of approach takes correspondences with more than two atomic terms into account,
but requires the ontologies to include matchable instances – a prerequisite that is not
fulfilled in many application scenarios. The approach proposed in this paper does not
require the existence of instances. Moreover, we can find related work in the field of
database schema matching. In [2] the authors describe complex matching as the task of
1 In the ODP taxonomy of patterns the notion of an alignment pattern is used instead of the
notion of a correspondence pattern. In particular, correspondence patterns are considered as
more general, having alignment patterns and reengineering patterns as subcategories. However,
we stick to the terminology introduced in [6] where an alignment is a set of correspondences.2 These patterns are now being included within OntologyDesignPatterns.org (ODP).3 http://alignapi.gforge.inria.fr/edoal.html
4
finding corresponding composite attributes (e.g., a name is equivalent with concatena-
tion of a first-name and a last-name). There are several systems dealing with this kind
of database schema matching (e.g., [1]).
According to Euzenat and Shvaiko [6] linguistic approaches to ontology matching
can be distinguished into language-based methods and methods which are based on
linguistic resources, whereas the more general class of terminological approaches also
includes string-based methods. The latter type of approach, i.e., similarity measures on
the lexical layer of ontologies, is part of almost every state-of-the-art matcher. There
is also a large body of work acknowledging the benefits of linguistic resources such
as WordNet when it comes to detecting lexical-semantic relations between concept or
property labels (see [7] for an overview). In addition, the low coverage of WordNet in
certain application domains has motivated the development of methods which leverage
more implicit evidence for those relationships [15], and of methods based on distri-
butional similarities which can be computed from textual information associated with
ontology entities [10, 8]. Only very few matchers, however, make use of natural lan-
guage processing techniques that go beyond tokenization and lemmatization [17].This
paper highlights the potential that lies within linguistic and in particular language-based
methods for ontology alignment.
4 Linguistic Analysis
In order to facilitate the integration of state-of-the-art tools and resources for natural
language processing into the matcher, we developed LiLA4 (Linguistic Label Analysis)
– a Java-based framework for the linguistic analysis of class and property labels which
provides a single uniform interface to the following open-source tools:
JWNL (version 1.4.1)5 is a programing interface for accessing the WordNet dictionary
(version 3.0)6 which contains information about more than 200,000 English words
and their lexical semantic relationships.
OpenNLP (version 1.3.0)7 is a framework for linguistic analysis including, for in-
stance, components for determining the lexical categories of words (e.g., adjective).
MorphAdorner (version 1.0)8 is a text processing framework which amongst other
components provides means for morphological analysis and generation, i.e., inflec-
tion of words.
LexParser (version 1.6.3)9 also known as the Stanford Parser is a syntactic parser
which can be used to determine the grammatical structure of phrases (e.g., noun
phrases such as “accepted paper”) or sentences.
4 http://code.google.com/p/lila-project/5 http://sourceforge.net/projects/jwordnet/6 http://wordnet.princeton.edu7 http://opennlp.sourceforge.net8 http://morphadorner.northwestern.edu9 http://nlp.stanford.edu/software/lex-parser.shtml
5
In addition, LiLA features a simple word sense disambiguation component and a
spell checker. The remainder of this section illustrates the core functionalities of LiLA
by virtue of a noun phrase, which serves as a running example.
paper written by clever students
Part-of-Speech Tagging. Each word in natural language belongs to a syntactic cat-
egory (or part-of-speech), which defines its basic syntactic behavior. Accordingly, a
part-of-speech tagger is a linguistic processing component for assigning appropriate
syntactic categories to a given set of words. While in principle, each POS tagger may
use its own set of category labels (tags), tag sets such as the Penn Treebank Tag Set10
for English are widely used, and certain naming conventions have emerged as quasi-
standards. Here, NN and NNS denote common nouns (singular or plural, respectively),
IN stands for a preposition, JJ indicates an adjective and VBN is the tag for a past
participle verb.
paper [NN] written [VBN] by [IN] clever [JJ] students [NNS]
Morphological Analysis. The field of morphology is concerned with the internal
structure of words, more precisely the morphological rules for inflection and word-
formation that enable humans to build a rich vocabulary from a basic inventory of mor-phemes – the smallest units in natural language that carry meaning. Each word consists
of one or more morphemes. Words that are built from more than one morpheme can
be split into a stem and an affix, i.e., a morph attached to a stem like “student”+“s”, for
example. In this case, the plural “s” is an inflectional morpheme, which alters the base
form (also called lexeme) without changing its syntactic category. A component which
reduces each word to its lemma (i.e., the canonical form of a lexeme which is typically
included in the lexicon of a language) is called a lemmatizer.
LiLA relies upon the MorphAdorner framework for performing both lemmatization
and morphological synthesis, i.e., the generation of specific word forms (e.g., “stu-
dents”) from lexemes (e.g., “student”). This also works well for irregular verbs such as
“write” for which we are able to generate, for instance, the past participle (“written”)
by means of conjugation. This way LiLA can convert between singular and plural of
the same noun (declination), as well as between active and passive voice or different
tenses of a given verb. The functionality to obtain derivations such as nominalizationsof verbs (e.g., “accept” and “acceptance”), for example, is provided by JWPL.
Lexical Semantic Analysis. Lexical semantics is the branch of linguistics that stud-
ies the meaning of words and their relationships. The popular lexical database of Word-
Net, which covers a wide range of such lexical semantic relations, is queried by LiLA
through the API of JWPL. Thus, given a word such as “clever” or “student”, LiLA
can get access to detailed information about the possible meanings of the word (see
homonymy), as well as about its synonyms, hyponyms, antonyms and otherwise related
senses (e.g., meronyms).
Synonymy, at least true synonymy, is rarely found in natural language. However, there
are many so-called near-synonyms (or plesionyms), i.e., words that share a common
10 http://www.cis.upenn.edu/ treebank/
6
meaning in a given context. Hence, two words are considered synonyms (or near-
synonyms) if they can be exchanged for one another in a sentence without altering
its truth conditions (e.g., “student” and “scholar”).
Homonymy and polysemy are types of semantic ambiguity. Two words are consid-
ered homonymous if they are spelled (homograph) and pronounced (homophone) in
the same way, while having distinct meanings (or senses). Homonyms with related
meanings, are called (regular) polysemes (e.g., “paper” as a substance or a writing
sheet made thereof). In case a query posed to the WordNet API returns multiple
senses for a given concept or property label, LiLA’s word sense disambiguationcomponent selects the most likely sense based on a vector-based representation of
the current lexical context.11
Hyponymy is a kind of subordination relating one lexical unit to another one with a
more general sense. The former is then called a hyponym, whereas the latter, i.e.,
the superordinate, represents the hypernym. Like meronymy, hyponymy is only
transitive within one and the same category (e.g., functional). A verb which is more
specific than another verb is sometimes called troponym (e.g., “write” and “cre-
ate”).
Antonymy is a kind of oppositeness that mostly holds between adjectives, but also
some verbs and even nouns can be considered antonyms if they exhibit opposite
semantic traits. One can distinguish between different types of antonyms such as
gradable (“early” and “late”), complementary (“acceptable” and “unacceptable”)
and relational (“student” and “professor”) antonyms.
Syntactic Parsing in computational linguistics typically refers to the analysis of
syntactic structures. Each parsing algorithm relies upon a certain grammar, that is a
formalism developed to describe the syntactically well-formed structures in a given
language. Essentially, two types of grammars – dependency grammars and phrase struc-
ture grammars – have emerged as the most wide-spread means to analyze and generate
syntactically well-formed utterances. The phrase structure depicted further below (left
column) has been generated by the Stanford Parser.12 NP and VP are phrasal categories
denoting a noun phrase or verb phrase, respectively.
(S(NP (NN paper))(VP (VBN written)(PP (IN by)(NP (JJ clever) (NNS students)))))
nsubj(written-2, paper-1)pobj(by-3, students-5)amod(students-5, clever-4)prep(written-2, by-3)
Given such a syntactic analysis and an appropriate set of rules for the English lan-
guage, we can determine (e.g., by means of OpenNLP or the Stanford Parser) the headof a phrase.13 In this case the head, i.e., the word which determines the category of a
phrase and carries the most essential semantic information, is the noun “paper”. Note
that the widely used heuristic of considering the right-most word as the head a phrase
11 For the experiments reported in this paper, we initialized this vector by adding all of the entity
labels found in the respective ontology.12 The phrasal category of the top-most node in the syntax tree should be NP (noun phrase) rather
than S, which denotes a sentence.13 We omit the distinction between syntactic and semantic heads.
7
(righthand head rule) only works well for morphological units such as compounds (e.g.,
“student paper”).
In addition, the Stanford Parser provides us with a list of syntactic dependenciesbetween the individual words of the phrase (right column). For example, it identifies
“paper” as the subject of “written” and “clever” as an adjective modifier of “students”.
In the following, we will explain how a linguistic analysis along the dimensions
outlined in this section can improve the results of a complex matching approach.
5 Matching Conditions
In the following we show how to use the linguistic analysis combined with a set of
simple structural techniques to detect complex correspondences. In particular, we
specify four correspondence patterns and define for each of them a set of matching
conditions. If each of these conditions is fulfilled we generate a correspondence as
instance of this pattern.
Class by Attribute Type (CAT) A correspondence A#1 ≡ ∃R#2 .B#2 of the CAT
pattern is generated by our algorithm, if the following conditions hold.
1. The label of B#2 is the nominalization of the modifier of the label of A#1 .
2. The class B#2 is subclass of the range of R#2 .
3. One of the following two conditions holds:
(a) The class A#1 is a subclass of the domain of R#2 due to the anchor alignment.
(b) The label of A#1 is a hyponym of the label of the domain of R#2 .
Fig. 2. Conditions relevant for detecting CAT correspondence Accepted Paper#1 ≡∃hasDecision#2 .Acceptance#2 .
A typical example is Accepted Paper#1 ≡ ∃hasDecision#2 .Acceptance#2 . In
Figure 2 we depict the three matching conditions relevant for this example: 1) The lin-
guistic analysis reveals that “Acceptance” is the nominalization of the active form of
“Accepted”, which is in turn the modifier of “AcceptedPaper”. A morphological analy-
sis indicates that the first condition is fulfilled. 2) We use a reasoner to check whether
Acceptance#2 is a subclass of the range of hasDecision#2 , and find that the second
condition is fulfilled, too. 3) The third condition is a disjunction. In this concrete case
the anchor alignment contains correspondence Paper#1 ≡ Paper#2 , which allows
us to conclude that the third condition is fulfilled. The third condition is a disjunction,
8
because on the one hand it might happen that the anchor alignment does not contain the
required correspondence, but the linguistic analysis detects the lexical-semantic relation
of hyponomy. On the other hand the linguistic analysis might fail, but the information
encoded in the anchor alignment might be sufficient. We defined similar disjunctions
for some of the other patterns.
In our previous approach we computed e.g., the edit-distance between “Acceptance”
and “Accepted” to detect a relation between AcceptedPaper#1 and Acceptance#2 . In
case it exceeded a certain threshold the counterpart of the first condition was fulfilled.
Similarity-based conditions are now replaced by conditions based on linguistic analysis.
Class by Inverse Attribute Type (CAT−1) A correspondence A#1 ≡ ∃R−1#2 . of the
CAT−1 type is generated if the following conditions hold.
1. The label of A#1 is the nominalization of the active form of the label of R#2 .
2. There exists a class B#2 which is a proper subclass of the range of R#2 .
3. One of the following two conditions holds:
(a) A#1 is, due to the anchor alignment, a subclass of B#2 .
(b) The label of A#1 is a hyponym of the label of B#2 .
This pattern and the conditions to detect it are similar to the CAT pattern and its
conditions. Due to the lack of space we omit a detailed description.
Class by Attribute Value (CAV) Here, we restrain ourselves to detect the boolean vari-
ant of the general CAV pattern whereby the the attribute values are true and false. Let
in the following adjm(X) refer to the adjective modifier of the phrase X , let advm(X)
be the adverbial modifier in X and let vp(X) refer to a verb phrase contained in X . A
correspondence A#1 ≡ ∃R#2 .{false} is generated by our algorithm, if the following
conditions hold.
1. The range of the datatype property R#2 is Boolean.
2. One of the following two conditions holds:
(a) The class A#1 is a subclass of the domain of R#2 due to the anchor alignment.
(b) The label of A#1 is a hyponym of the label of the domain R#2 .
3. advm(label(R#2 )) is the antonym of advm(adjm(label(A#1 ))).
4. The head of label(R#2 ) is the nominalization of vp(adjm(label(A#1 ))).
Fig. 3. Conditions relevant for detecting CAV correspondence
Late − Registered Participant#1 ≡ ∃earlyRegistration#2 .{false}.
9
Regarding this pattern we use the linguistic analysis to detect antonyms. We
expect that modifiers, which are antonyms, will be used to describe a pair of disjoint
classes. Complex expressions as ∃R#2 .{true} and ∃R#2 .{false}, given that R#2 is a
functional property, refer also – for logical reasons – to a pair of disjoint classes.
Inverse Property (IP) A correspondence R−1#1 ⊆ P#2 of type IP is generated, if all
following conditions hold.
1. The verb phrase of the label of R#1 is the active voice of the verb phrase of the
label of P#2 .
2. One of the following two conditions holds:
(a) The domain of R#1 is a subclass of the range of P#2 .
(b) The label of the domain of R#1 is a hyponym of the label of the range of P#2 .
3. One of the following two conditions holds:
(a) The range of R#1 is a subclass of the domain of P#2 .
(b) The label of the range of R#1 is a hyponym of the label of the domain of P#2 .
The IP pattern is the simplest pattern regarding its set of conditions. The first con-
dition is based on the fact that two properties are inverse properties with higher proba-
bility, if both contain the same verb phrase in a different voice (active or passive voice).
The two other structural conditions ensure that there is a subsumption (possibly equiv-
alence) relation between domain and range of both properties.
It is surprising that it is sometimes harder to detect a simple property equivalence or
subsumption than an instance of the IP pattern. An example found in our experiments
is the following one. In one ontology we have the property writtenBy#1 and its inverse
authorOf #1 , while in the other ontology we have a property write paper#2 . Regard-
ing these properties there are two correct correspondences, namely authorOf #1 ⊆write paper#2 and writtenBy−1
#1 ⊆ write paper#2 . While the first one is hard to de-
tect, the second one fulfills all of the conditions listed above and is thus detected by our
approach. It is now possible to derive the first correspondence from the second one.
6 Experiments
For our experiments we used the dataset of the OAEI conference track [16]. This dataset
consists of several, relatively expressive ontologies that describe the domain of orga-
nizing conferences from different perspectives. Regarding this dataset, we made exper-
iments on the full set of ontologies.14 However, since we want to compare our approach
to its predecessor, we also present results restricted to a subset of 9 ontologies that we
have used in [11]. To decide whether the structural conditions hold, we used the Pellet
reasoner [13]. For checking the linguistic conditions we rely on LiLA. Structural con-
ditions that compare entities from different ontologies require an anchor alignment of
non-complex correspondences. We generated this alignment by computing a similarity
measure based on the Levensthein distance thresholding the results at a high value.
14 Since we could not process the ontologies LINKLINGS, COCUS, CONFIOUS with the Pellet
reasoner we have excluded them from our experiments.
10
Dataset Approach True Positives False Positives Precision
CAT CAT−1 CAV IP∑
CAT CAT−1 CAV IP∑
subset similarity 4 3 2 - 9 4 7 0 - 11 0.450
subset linguistic 4 2 2 8 16 1 0 0 0 1 0.941
full set linguistic 9 4 2 17 32 5 1 0 1 7 0.821
Table 1. Experimental results in terms of true and false positives.
In Table 1, the approach described in this paper is referred to as linguistic approach,
its predecessor is called similarity approach. For each of the four patterns we show
the number of true and false positives, as well as the precision of the approach. We
can see that the use of linguistic methods helped us to increase precision of the overall
approach by a large degree from 45% to 94%, while the number of correctly detected
correspondences stays nearly stable. Notice that based on the similarity approach it was
not possible to define matching conditions for the IP pattern. If we include the IP pattern
in our analysis, we come up with the conclusion that we signficantly increased recall.
With the similarity approach we could additionally define conditions for a pattern
called property chain (not depicted here). We omitted this pattern here, as we thought
that the rationales underlying this pattern are hard to justify and that the chosen con-
ditions were partially overfitting to certain aspects of the dataset. Note that the values
given for the similarity approach are based on the optimal threshold: raising or lowering
the threshold results in a clear loss of recall or precision, while the additional gain is
rather limited as shown in [11]. The linguistic methods are different in that there is no
threshold whose value would be crucial for the overall performance of the approach.
Our previous approach has been criticized for the requirement of a correct and com-
plete input alignment. However, the new results indicate that such an input alignment is
not necessary. It is sufficient to generate an incomplete and partially incorrect alignment
in a prior step and to use it as an anchor in the subsequent matching process. Thus, the
approach is robust against the noise introduced by such an imperfect anchor alignment.
7 Conclusion
In this paper we have described how to integrate linguistic techniques into a pattern-
based approach for detecting complex correspondences. In particular, we have pre-
sented correspondence patterns and defined for each of them a set of matching con-
ditions. While in a previous approach [11] many of these conditions were based on
computing a simple string-based similarity value, we argued now that it is more appro-
priate to substitute these conditions by a set of conditions that make use of a linguistic
analysis. In our experiments we showed that the new approach yields a significantly
higher precision. The tool used to conduct the experiments is open source and available
online.15 Due to its modular structure, matching conditions for new correspondence
pattern can easily be specified in a generic XML syntax.
Acknowledgments Johanna Volker is financed by a Margarete-von-Wrangell scholar-
ship of the European Social Fund (ESF) and the Ministry of Science, Research and
the Arts Baden-Wurttemberg. Ondrej Svab-Zamazal is partly supported by grant no.
P202/10/1825 of the Grant Agency of the Czech Republic.
15 http://code.google.com/p/generatingcomplexalignments/
11
References
1. Robin Dhamankar, Yoonkyong Lee, Anhai Doan, Alon Halevy, and Pedro Domingos. iMAP:
discovering complex semantic matches between database schemas. In Proceedings of theACM SIGMOD International Conference on Management of Data, 2004.
2. AnHai Doan and Alon Y. Halevy. Semantic-integration research in the database community.
AI Magazine, 26:83–94, 2005.
3. Jerome Euzenat. An API for ontology alignment. In Proceedings of the 3rd InternationalSemantic Web Conference (ISWC), pages 698–712, 2004.
4. Jerome Euzenat, Alfio Ferrara, Laura Hollink, Antoine Isaac, Cliff Joslyn, Veronique
Malaise, Christian Meilicke, Andriy Nikolov, Juan Pane, Marta Sabou, Francois Scharffe,
Pavel Shvaiko, Vassilis Spiliopoulos, Heiner Stuckenschmidt, Ondrej Svab-Zamazal, Vojtech
Svatek, Cassia Trojahn dos Santos, George A. Vouros, and Shenghui Wang. Results of the
ontology alignment evaluation initiative 2009. In Proceedings of the 4th International Work-shop on Ontology Matching (OM-2009), 2009.
5. Jerome Euzenat, Francois Scharffe, and Antoine Zimmermann. Expressive alignment lan-
guage and implementation. Deliverable 2.2.10, Knowledge web, 2007.
6. Jerome Euzenat and Pavel Shvaiko. Ontology Matching. Springer, 2007.
7. Feiyu Lin and Kurt Sandkuhl. A survey of exploiting WordNet in ontology matching. In Max
Bramer, editor, IFIP AI, volume 276 of IFIP, pages 341–350. Springer, September 2008.
8. Giuseppe Pirro and Domenico Talia. LOM: a linguistic ontology matcher based on informa-
tion retrieval. Journal on Information Science, 34(6):845–860, 2008.
9. Han Qin, Dejing Dou, and Paea LePendu. Discovering Executable Semantic Mappings
Between Ontologies. On the Move to Meaningful Internet Systems 2007: CoopIS, DOA,ODBASE, GADA, and IS, pages 832–849, 2007.
10. Yuzhong Qu, Wei Hu, and Gong Cheng. Constructing virtual documents for ontology match-
ing. In Proceedings of the 15th International Conference on World Wide Web (WWW), pages
23–31, New York, NY, USA, 2006. ACM.
11. Dominique Ritze, Christian Meilicke, Ondrej Svab-Zamazal, and Heiner Stuckenschmidt.
A pattern-based ontology matching approach for detecting complex correspondences. In
Proceedings of the ISWC workshop on ontology matching, Washington DC, USA, 2009.
12. Francois Scharffe. Correspondence Patterns Representation. PhD thesis, University of Inns-
bruck, 2009.
13. Evren Sirin, Bijan Parsia, Bernardo Cuenca Grau, Aditya Kalyanpur, and Yarden Katz. Pel-
let: a practical OWL-DL reasoner,. Journal of Web Semantics, 5:51–53, 2007.
14. Ondrej Svab-Zamazal, Vojtech Svatek, and Luigi Iannone. Pattern-based ontology transfor-
mation service exploiting OPPL and OWL-API. In Knowledge Engineering and KnowledgeManagement by the Masses. EKAW-2010., 2010. Accepted.
15. Willem Robert van Hage, Sophia Katrenko, and Guus Schreiber. A method to combine lin-
guistic ontology-mapping techniques. In Yolanda Gil, Enrico Motta, V. Richard Benjamins,
and Mark A. Musen, editors, International Semantic Web Conference (ISWC), volume 3729
of LNCS, pages 732–744. Springer, November 2005.
16. Ondrej Svab, Vojtech Svatek, Petr Berka, Dusan Rak, and Petr Tomasek. Ontofarm: To-
wards an experimental collection of parallel ontologies. In Poster Track of the InternationalSemantic Web Conference (ISWC), Galway, Ireland, 2005.
17. Pinar Wennerberg, Manuel Moller, and Sonja Zillner. A linguistic approach to aligning repre-
sentations of human anatomy and radiology. In Proceedings of the International Conferenceon Biomedical Ontologies (ICBO), 7 2009.
12
Lost in Translation? Empirical Analysis of MappingCompositions for Large Ontologies
Anna Tordai,1 Amir Ghazvinian,2 Jacco van Ossenbruggen,1 Mark A. Musen,2
Natalya F. Noy2
1 VU University Amsterdam, Amsterdam, The Netherlands
{atordai,jrvosse}@few.vu.nl2 Stanford University, Stanford, CA 94305, US
{amirg,musen,noy}@stanford.edu
Abstract. When three or more ontologies have been aligned, longer chains of
mapped concepts start to appear. In this paper, we empirically study the nature
of these composite mappings, focusing on chains of (near) equivalence links
of length two. We ask human experts to evaluate samples of composite map-
pings, taken from large real life data sets. Based on these evaluations, we analyze
the features of mappings produced by composition in three different domains
(bio-medicine, cultural heritage, and library subject headings), among ontologies
in multiple languages (English, Dutch, German, and French), and using exist-
ing mappings that were created by different methods (lexical and instance-based
methods). We examine the quality of the composite mappings relative to the qual-
ity of the input mappings and analyze how characteristics of the input mappings
and the ontologies influence the composition.
1 Introduction
Researchers typically study ontology alignments in the context of a single source and
target ontology. As more and more of such alignments are being created and published,
however, longer chains of equivalent or otherwise related concepts start to emerge in
our data sets. In this paper, we analyze the quality of a subset of such chains, focusing
on short chains of equivalence and near equivalence links. Most of us have clear intu-
itions about the properties of such chains. For example, equivalence relations such as
owl:sameAs and skos:exactMatch, are defined as being transitive, so it should
be safe to assume that if term A is equivalent to B, and B is equivalent to C, then Ashould also be equivalent to C. We will test this hypothesis empirically by determin-
ing to what extent such transitivity actually holds in our data sets, and if not, what is
going wrong Furthermore, for relations such as skos:closeMatch, which are not
defined as being transitive, we might ask how often chains of these relations turn out to
be transitive after all.
We use the notion of a mapping as defined in [15]. Given a mapping from A to B
and from B to C, where concepts A, B and C are part of three different ontologies, we
call the mapping from A to C a composite mapping. Although mapping composition
is related to the use as background knowledge where concept B would be part of the
background ontology [2], we do not predefine ontologies as a source of background
13
knowledge. We analyze the properties of such composite mappings on real life data
sets, addressing the following two research questions:
– What is the quality of composite mappings relative to the quality of input map-
pings?
– Does the quality of composite mappings depend on other characteristics of the input
mappings or ontologies?
In order to answer these research questions, we study composite mappings for on-
tologies in different domains, using input mappings generated in different ways (Sec-
tion 3.1). We analyzed the precision of composite mappings by sampling them and
having human experts verify the samples (Section 3.3). In some cases, we already had
pre-existing alignments for the sets of ontologies for which we analyze composite map-
pings. In these cases, we compared the precision of the composed mappings with the
precision of existing mappings. We then analyzed our results (Section 5) and made ob-
servations regarding the quality and quantity of composed mappings, trying to identify
reasons for correct and incorrect mapping compositions based on characteristics of the
data and the input mappings.
The main contribution of this paper is a large-scale empirical analysis of the nature
of composite mappings given varied sets of input ontologies and mappings.
2 Related Work
Researchers in the area of database schema matching have studied mapping composi-
tion extensively [1, 14, 4]. However, these researchers have focused on the use of map-
ping composition to perform query transformation rather than for generating mappings.
In ontology matching, Euzenat [6] discusses mapping composition in a theoretical
paper on algebras of relations as a means for validating existing mappings and creating
new mappings. This work considers composition through equivalence mappings to be
a trivial case because the result is an equivalence relation, and because we can assume
that equivalence is transitive. In practice, however, automatically generated mappings
are usually similarity mappings at best, and therefore the composition of such mappings
is not trivial. We look at such automatically generated mappings and analyze results of
composition to find out whether they are interesting or truly lost in translation.
Researchers have already developed a plethora of tools for generating mappings
and compared their performance at the OAEI. These off-the-shelf tools, such as AS-
MOV [12], RiMOM [22], Falcon-AO [11], and DSSim [16] perform well on OAEI
benchmarks and on certain specialized tracks. However, the results of the 2009 library
track showed that current tools largely fail on extremely large vocabularies and vocab-
ularies that use multiple languages [7].
Mapping composition has some parallels to the use of background knowledge by
mapping tools. Tools such as SAMBO [13] and ASMOV use background knowledge
(UMLS Metathesaurus, WordNet) to improve the quality of mappings. When mapping
two domain ontologies, these tools either use existing mappings from these domain
ontologies to some background source, such as UMLS or WordNet, or create these
mappings “on the fly” through lexical comparison or other means. The tools then use
14
Set Domain Ontologies Language Ontology size
BioPortal Biomedicine 151 ontologies English Ranging from under
from BioPortal 100 concepts to 380K concepts
Mean size=17,805 (SD= 61,614)
Total concepts: 2,688,609
CH Cultural Her-
itage
Getty’s Art and Ar-
chitecture Thesaurus
(AAT)
English and
Dutch
27,077 concepts with English and
Dutch labels
Princeton WordNet English 115,424 synsets with
203,147 English labels
Cornetto Dutch 70,370 synsets and
103,762 Dutch labels
Library General Library of Congress
Subject Headings
(LCSH)
English 339,612 concepts
Rameau French 157,287 concepts
SWD German 163,175 concepts
Table 1. Sets of ontologies that we used in mapping composition and their characteristics.
these mappings to a single source of background knowledge for creating mappings for
the domain ontologies. This method is related to mapping composition because we use
a mapping to a third ontology or vocabulary. In this sense, in mapping composition anyontology becomes a source of background knowledge.
The COMA [5] and COMA++ [3] tools combine several matching techniques in-
cluding composition of mappings. The evaluation of the tools demonstrated the effec-
tiveness of mapping composition without going into a more detailed analysis of the
results.
3 Materials and Methods
In this section, we describe the ontologies and existing mappings that we used for map-
ping composition (Section 3.1), the method for creating compositions and its complex-
ity (Section 3.2), and our methodology for assessing the precision of the composed
mappings (Section 3.3).
3.1 Data: Ontologies and Input Mappings
In order to get a comprehensive analysis of mapping composition under different condi-
tions, we considered three sets of ontologies and mappings. We have ontologies in three
different domains: biomedicine, cultural heritage and library subject headings (Table 1).
The terms in these ontologies have labels in four languages: English, Dutch, German
and French, and the input mappings we use for composition were generated using two
types of methods: lexical method, and instance-based method (Table 2).
15
Set Method for mapping Number of mappings used Precision
generation for composition
BioPortal Lexical comparison of 575,642 mappings 0.94
preferred names and 459,941 Preferred–Preferred 0.99
synonyms 115,701 Preferred–Synonym 0.76
CH Lexical comparison 6,914 AAT–Cornetto 0.88
of labels 4,592 AAT–WordNet 0.82
3,144 Cornetto-WordNet 0.95
Library Instance-based 2,242 LCSH–Rameau 0.95
2,334 SWD–LCSH 0.54
685 Rameau–SWD 0.72
Table 2. Input mappings that we used for mapping composition for the three sets of ontologies.
Our first set of ontologies came from BioPortal [17], a Web-based repository of
biomedical ontologies. At the time we collected the data, BioPortal contained 151 on-
tologies with more than 2.5 million concepts among them. We generated mappings
between these ontologies using simple lexical comparisons of preferred names and syn-
onyms after normalization [9, 8].
The second set of mappings links three large vocabularies in the cultural-heritage
domain: Getty’s Art and Architecture Thesaurus (AAT3, extended with Dutch labels
from AATNed4), Princeton WordNet5 version 2.0 and Cornetto,6 a WordNet-like lexical
resource for Dutch. We generated mappings between AAT and WordNet, and between
AAT and Cornetto using simple lexical comparison [19]. The Cornetto project [20]
created mappings between Cornetto and different versions of WordNet using a combi-
nation of manual and automatic methods.
Finally, we used a set of ontologies and mappings from the Library track in the
OAEI 2009. This set contains three lists of subject headings for describing content of
books: the Library of Congress Subject Headings (LCSH); Rameau, a list used by the
French National Library; and the Subject Heading Authority File (SWD), which is used
by the German National Library. Each list contains from 150,000 to 300,000 concepts.
We used the mappings that Wang and colleagues [21] created using instance-based
matching based on books that were classified using terms from more than one vocab-
ulary. This method for generating mappings ranks the resulting mappings according to
confidence level. Although there are a total of almost 2 million mappings, over 90%
of them have confidence measure lower than 0.1. For the purpose of composing map-
pings, we selected only those mappings that had a confidence measure greater than 0.7.
3 http://www.getty.edu/research/conducting_research/vocabularies/aat/
4 http://www.aat-ned.nl/5 http://wordnet.princeton.edu/6 http://www2.let.vu.nl/oz/cornetto/index.html
16
We estimate the precision of these mappings by evaluating samples manually. These
mappings involve fewer than 1.5% of the concepts in the vocabularies.
In the cultural heritage and OAEI library track the number of input mappings is
significantly lower than in the BioPortal case, as our aim was to select high-quality
mappings. We chose a representative subset in order to analyze the properties of map-
ping composition.
3.2 Computing Mapping Composition
In this paper, we consider only composition of two mappings. The BioPortal compo-
sitions were computed using a relational database, and the cultural heritage and OAEI
library track composition algorithms were written in SWI-Prolog.7.
Since we had detailed information on the source of the input mappings for all Bio-
Portal ontologies, we further analyzed the composed mappings for BioPortal to under-
stand better how characteristics of input mappings affect the number and precision of
composed mappings. To perform such analysis, we broke down the composed map-
pings into groups, based on the types of input mappings that contributed to the com-
position. There are two types of input mappings (see Table 2): Preferred–Preferred and
Preferred–Synonym mappings. We do not include Synonym–Synonym mappings in our
input because they have low precision(0.36). Different combinations of the input map-
pings produce six groups of composed mappings which are displayed in Figure 1.
For instance, Figure 1A illustrates the case where we composes a mapping from a
preferred name for the concept C1 to a preferred name for C2 with a mapping from
the preferred name for C2 to the preferred name of C3. We refer to this case as PPP .
Note that this composition produces a subset of the Preferred–Preferred mappings be-
tween O1 and O3. PSP mappings (Figure 1B) also produce a subset of the Preferred–
Preferred mappings. Similarly, PPS mappings (Figure 1C) and SPS mappings (Fig-
ure 1D) produce subsets of the Preferred–Synonym and Synonym–Synonym mappings
between O1 and O3, respectively. We analyze these subsets and compare their preci-
sions to those of the original Preferred–Synonym and Synonym–Synonym mappings
that were generated directly by comparing O1 and O3. Figure 1E and F illustrate the
other two cases, PSPS and PSSP , which produce mappings that we cannot obtain by
comparing preferred names and synonyms directly.
3.3 Sampling and evaluation
In order to evaluate the precision of the composed mappings as well as the precision
of input mappings (see Table 2), we sampled the mappings and evaluated the samples
manually. Because of the scale of our data—with hundreds of thousands of mappings to
verify—evaluating all the mappings manually was not feasible. Furthermore, because of
the size of the ontologies themselves, creating a complete set of mappings so that we can
evaluate recall was not feasible either. In addition, the recall of mapping composition is
necessarily limited by the recall of the input mappings used for composition. Thus, we
focus in this evaluation on estimating only the precision of the composed mappings.
7 http://www.swi-prolog.org/
17
For BioPortal mappings, we used stratified sampling [10] to select mappings for
manual evaluation. Among the BioPortal ontologies, there is a large number of ontology
pairs that have only one or two composed mappings between them. At the same time,
there are pairs of ontologies that have thousands of mappings between them. Therefore,
we constructed the strata to ensure that our samples include mappings between ontology
pairs with only a few mappings between them, as well as mappings between ontology
pairs with thousands of mappings, and clusters in between. We sampled a total of 2350
mappings from the different BioPortal mappings sets. Our sample sizes ranged from
210 to 400 mappings per set depending on the number of original mappings.
In the case studies involving cultural heritage and library subject headings, we eval-
uated manually all mapping sets containing fewer than 500 mappings and took samples
of 100 mappings from larger sets. We sampled the total of approximately 1,000 map-
pings from these sets.
Human experts evaluated the samples using the evaluation tool used in [18] for the
cultural heritage and Library track data, and a similar tool for the BioPortal data, and
categorized each mapping into one of six categories: exact match, close match, broadermatch, narrower match, related match, or incorrect. For measuring precision, we con-
sidered only exact and close matches as correct. A detailed analysis of the broader,
narrower and related matches is out of scope for this paper but we plan to perform it
in the future. We measured agreement using Cohen’s kappa on subsets of samples be-
tween raters, finding substantial agreement for BioPortal (0.72) and cultural heritage
evaluation (0.70) and almost perfect agreement with the manually evaluated mappings
used in the OAEI library track (0.85).
4 ResultsIn this section, we present the precision of mapping composition for the three sets of
ontologies in our study. We discuss these results in Section 5.
4.1 Results: Biomedical Ontologies
Figure 2A shows the results for the overall precision of composed mappings. Using
575,642 input mappings with precision 0.94, we generated 599,625 composed map-
A CB
C1 C2 C3
P PPS SS
SPS Mapping
C1 C2 C3
PSSP Mapping
C1 C2 C3
P PPS SS
PSPS Mapping
C1 C2 C3
P PPS SS
PSP Mapping
C1 C2 C3
P PPS SS
PPS Mapping
C1 C2 C3
P PPS SS
PPP MappingD FE
P PPS SS
S
Fig. 1. Methods for composing mappings between concepts in three different ontologies (C1 ∈O1, C2 ∈ O2, C3 ∈ O3) using mappings between preferred labels (P ) and synonyms (S).
Figure A illustrates the PPP mappings: a composition of a mapping from a preferred name of
C1 to preferred name of C2 with the mapping between preferred names of C2 and C3. Figure B
illustrates PSP mappings: a match of C1 preferred name to C2 synonym with a match of C2
synonym to C3 preferred name. Figures C-F illustrate the remaining possible cases.
18
pings with a precision of 0.92. Figures 2B, 2C, and 2D show the precision of composi-
tion for different cases from Figure 1. We group these cases by the sets of input map-
pings that they used. Composing Preferred–Synonym mappings, which had a precision
of 0.76, yielded 147,438 composed mappings with precision 0.84. Other combinations
(Figures 2C and 2D) resulted in sets of composed mappings with precisions similar to
the precisions of the input mappings.
Figure 3 provides additional information on the precision of the individual cases.
The two cases that resulted in the subset of what we could have obtained directly
by comparing preferred names lexically (PPP and PSP ), provided mappings with
the highest precision, 0.99. The SPS mappings constitute a subset of the Synonym–
Synonym mappings for O1 and O3. We did not use these types of mappings as input
mappings because they have very low precision, 0.36. However, using mapping compo-
sition to identify a subset of Synonym–Synonym mappings almost doubles the precision
of these mappings, bringing it up to 0.6.
Additionally, using composition, we identified mappings without lexical similar-
ity in their preferred names or synonyms (PSPS and PSSP mappings). Such map-
pings can be identified by composition through a concept with lexical similarity to
both mapped concepts. These two cases produced 50,353 new mappings with the pre-
cision of 0.68. For example, we found a PSSP mapping between the concept CRA-
NIAL SKELETON from the Amphibian gross anatomy ontology and SKULL from the
Foundational Model of Anatomy. These two concepts each map to the concept CRA-
NIUM from the Teleost anatomy and development ontology, which has the synonyms
CRANIAL SKELETON and SKULL.
4.2 Results: Cultural Heritage
Figure 4A shows the results of mapping composition for the cultural heritage domain.
The precision of composed mappings is at least 0.8 in all three cases, with the number
of mappings identified through composition ranging from 263 to 1,774. In fact, the
composed mappings between Cornetto and WordNet have a precision of 0.9.
Because we have lexical mappings available for this set, we can compare the com-
posed mappings to the lexical ones, and analyze how many non-lexical mappings we
generate by composing lexical mappings.
Upon closer examination of the mappings, we found that 134 (30%) of the com-
posed mappings between AAT and WordNet have little or no lexical similarity. For
example, through composition we mapped TOBACONNISTS’ SHOP to TOBACCO SHOP
and WATCHMEN to GUARD. Similarly, we found 110 non-lexical mappings between
AAT and Cornetto, such as BADKLEDING to BADKOSTUUM, both of which mean “bathing
suit” in Dutch. This subset of composed mappings not including lexical similarity has a
precision of 0.56, which is lower than the precision of composed mappings in general.
Between Cornetto and WordNet, 1,208 of the 1,774 composed mappings are listed
as “near equal synonym” mappings in the original Cornetto-WordNet mappings of the
Cornetto project. These are not the same as the “equal synonym” mappings used as
input mappings for other compositions. Another 448 of the composed mappings are
entirely new and have an average precision of 0.7.
19
Pr= 0.99(459,941)
Pr= 0.76(115,701)
Pr= 0.75(91,808)
O1 O2 O3Pr= 0.99(459,941)
Pr= 0.99(459,941)
Pr= 0.99(360,379)
P PPS SS P PP
S SS
O1 O2 O3
O1 O2 O3Pr= 0.76(115,701)
Pr= 0.84(147,438)
Pr= 0.76(115,701)
P PPS SS
P PPS SS
P PPS SS
S
P PPS SS
O1 O2 O3Pr= 0.94(575,642)
Pr= 0.92(599,625)
Pr= 0.94(575,642)
A B
C D
Fig. 2. Mapping composition results for BioPortal ontologies. O1, O2 and O3 represent any three
ontologies linked through mappings from Bioportal. Figure A (the shaded diagram) shows the
overall precision of the input mappings and their number in parentheses on the solid lines. It
shows the precision of composed mappings and their number above the dotted line. Figures B, C,
and D provide details for the precision of composed mappings, grouped by the precision of input
mappings. Figure B contains the mappings that used only Preferred–Synonym mappings as in-
put; Figure C contains the mappings that composed Preferred–Preferred mappings; and Figure D
provides the data for the composition of Preferred–Preferred mappings and Preferred–Synonym
mappings.
!"##$ !"##$
!"%&$ !"%&$ !"%’$!"($
!"##$ !"##$!"))$ !")%$ !")%$
!"%)$
!$
!"*$
!"+$
!"($
!")$
,$
-./01$234$15267$ 836.94:3;$<=/241=>$32==/?1=>$=1.2714$
P PPS SS
P PPS SS
P PPS SS
S
!"#$%&%’(
(360,379) (70,075) (91,808) (27,010) (32,863) (17,490)
P PPS SS
P PPS SS
P PPS SS
Fig. 3. Mapping composition results for BioPortal ontologies. The bar graph shows precision
for composed mappings. The (lighter) left bar shows precision of exact and close matches, the
(darker) right bar shows the precision if we include broader, narrower, and related matches. Num-
bers in parentheses indicate the total number of mappings
4.3 Results: The OAEI Library Track
Figure 4B shows the results of mapping composition using the library subject headings
mappings. Precision of the composed mappings is higher than 0.74 and the number
20
Cornetto WordNet
AAT
Pr= 0.82(437)
Pr= 0.80(263)
Pr= 0.90(1,774)
Pr= 0.82(4,592)
Pr= 0.95(3,144)
Pr= 0.88(6,914)
SWD Rameau
LCSH
Pr= 0.86(146)
Pr= 0.89(132)
Pr= 0.74(266)
Pr= 0.95(2,242)
Pr= 0.54(2,334)
Pr= 0.72(685)
A B
Fig. 4. Mapping composition results for cultural heritage domain (Figure A) and OAEI library
track (Figure B). The numbers in bold outside the triangle show the precision and the number
of composed mappings in parentheses. The numbers inside the triangle show the precision and
number of input mappings
!"#$ !"#%$!"&$ !"#&$ !"#’$
!"()$
!"&%$ !"#&$!"&*$ !"&&$
!"&+$!"#’$
!$
!"%$
!")$
!"’$
!"#$
,$
-./012.3445$ 44536 7$ -./012.36 7$ 86 93:-8;$ :-8;3<=> 1=?$ <=> 1=?386 9$
[email protected]$=0B$1C=DE$ F0D@?BG0H$I/.=B1/J$0=//.K1/J$/1@=E1B$!"#$%&%’(
(263) (437) (1,774) (132) (146) (266)
Fig. 5. Mapping composition results for cultural heritage and library-track ontologies. The bar
graph shows precision for composed mappings. The (lighter) left bar shows precision of exact
and close matches, the (darker) right bar shows the precision if we include broader, narrower, and
related matches. Numbers in parentheses indicate the total number of mappings
of generated mappings ranges from 132 between the Subject Heading Authority File
(SWD) and the Library of Congress Subject Headings (LCSH) and 266 between SWD
and Rameau (a list used by the French National Library).
In two cases—mappings between SWD and LCSH and mappings between Rameau
and SWD—the composed mappings actually had higher precision than the input map-
pings.
In this case, we also compared the composed mappings to the input mappings. We
found that, of the 132 mappings between SWD and LCSH, 13 (10%) mappings did not
overlap with any of the original instance-based mappings, including those that had a
confidence measure lower than 0.7. In other words, for these 13 mappings, there were
no instances (books) available. For LCSH and Rameau, we found 8 (5%) such “new”
mappings, and for Rameau and SWD, 65 (24%) mappings. The high number of new
composed mappings between Rameau and SWD is due to the low number of instances
available for creating the original mappings. However, the precision of these subsets is
lower: 0.37 between LCSH and Rameau, 0.54 between Rameau and SWD, and 0.92
between SWD and LCSH.
21
4.4 Broader, Narrower, and Related MappingsWhen evaluating the composed mappings, we have also recorded whether each mapping
represented a narrower, broader, or related mapping, rather than a close or exact match.
Figures 3 and 5 show the increase in precision of composed mappings if we also count
broader, narrower, and related mappings as correct. Figure 5 includes the data for both
the cultural-heritage and the library-track case. The increase in precision in both of
these cases is less dramatic than for the biomedical ontologies. In this case, the average
increase in precision was 11%, whereas for BioPortal ontologies the average increase
was 14%, with the most significant increase (30%) in the PSSP case.
5 DiscussionIn this paper, we have presented the results of our analysis of mapping composition in
three different domains. Our results show that the quality of composed mappings de-
pend on the ontology characteristics, and the content and quality of the input mappings.
The characteristics of the ontologies, such as the way they are implemented or the
way preferred labels and synonyms are used, have a profound effect on composition.
For example, in Medical Subject Headings (MeSH) concepts often have narrower terms
as synonyms. The concept TREMORS in MeSH has a synonym NERVE TREMOR, which
in reality is a narrower term, not a synonym. As a result, many of the composed map-
pings that involved MeSH terms were not close matches but rather broader or narrower
mappings.
It is clear that the number of input mappings determines the number of composed
mappings, but we see in our results that there are large variations in the number of input
mappings in three cases studies. This is partly due to the size of the ontologies and
partly because of the confidence level of the original input mappings which is a limiting
factor for example in the Library track case study.
The content of the ontologies also influences the quality of the mapping composi-
tions. When the content overlaps, meaning the domains of the ontologies are the same
or very similar, the meaning of the concepts is also closer, and the composed mappings
are likely to be equivalence mappings rather than broader, narrower or related map-
pings. In the cultural heritage case study Cornetto and WordNet are unlikely to cover
art and architectural concepts, reducing the chance of creating equivalence composi-
tions between AAT and Cornetto, and AAT between AAT and WordNet.
Finally, the quality of input mappings has a direct effect on the quality of mapping
compositions. High quality input mappings tend to result in high quality mapping com-
positions. Intuitively, the precision of the compositions should be the product of the
precisions of the input mappings. However, especially in the BioPortal data we find
cases where the precision of compositions exceeds the precision of input mappings
(Figure 3). We find similar cases in both the AAT and Library track case studies.
We also found that many of the composed mappings though not exact, or close
matches nevertheless represent a semantic relationship such as broader, narrower or
related (Figure 3 and 5). For example, the concept BLURRED VISION from the “Sug-
gested Ontology for Pharmacogenomics” maps to composition to VISION ABNORMAL
in the “WHO Adverse Reaction Terminology”, forming a narrower relationship be-
tween the two concepts. This kind of semantic drift between concepts seems to arise
22
often through mapping composition caused by ontology characteristics, or concepts de-
viating in meaning in different languages.
In our future work we plan to perform a more detailed evaluation of the content of
the mappings to determine why the precision of the composed mappings exceeds the
precision of the input mappings in certain cases. We also need to study the effect of
semantic drift by analyzing the relationship between the type of equivalence relation in
input mappings and compositions, and extend our scope to longer composition chains.
6 ConclusionIn this paper, we presented an empirical analysis of the quality of mapping composition
for various use cases. We conclude that mapping composition produced mappings of
comparable quality to the input mappings; precision of the composed mappings is not
much worse than the original precision of mappings, and sometimes it is even better.
Even when composing lexical mappings, in some cases we produced mappings that
lexical methods would not produce. Finally, the quality and the number of composed
mappings can be affected significantly by the characteristics of the ontologies them-
selves, the type of input mappings, the number and coverage of the input mappings.
Our results confirm our intuitions on mapping composition. The contribution of this pa-
per is that we have tested these intuitions empirically and validated them using methods
described in literature.
AcknowledgmentsMany thanks to Guus Schreiber and Bob Wielinga for their help and comments on this paper. We
thank Antoine Isaac, Shenghui Wang and the Telplus project for providing the library track data
and their mappings, and for the many explanations they provided. We also thank the Cornetto
project for providing data and mappings and Babette Heyer for helping with the evaluation of
the mappings. This work was supported in part by the National Center for Biomedical Ontology,
under roadmap-initiative grant U54 HG004028 from the National Institutes of Health, and partly
by the EC eContentplus programme via de EuropeanaConnect project.
References
1. K. Aberer, P. Cudre-Mauroux, and M. Hauswirth. A framework for semantic gossiping.
SIGMOD Rec., 31(4):48–53, 2002.
2. Z. Aleksovski, M. C. A. Klein, W. ten Kate, and F. van Harmelen. Matching unstructured
vocabularies using a background ontology. In EKAW, pages 182–197, 2006.
3. D. Aumueller, H.-H. Do, S. Massmann, and E. Rahm. Schema and ontology matching with
coma++. In SIGMOD ’05: Proceedings of the 2005 ACM SIGMOD international conferenceon Management of data, pages 906–908. ACM, 2005.
4. P. A. Bernstein, T. J. Green, S. Melnik, and A. Nash. Implementing mapping composition.
In VLDB ’06: Proceedings of the 32nd international conference on Very large data bases,
pages 55–66. VLDB Endowment, 2006.
5. H.-H. Do and E. Rahm. Coma: a system for flexible combination of schema matching ap-
proaches. In VLDB ’02: Proceedings of the 28th international conference on Very LargeData Bases, pages 610–621. VLDB Endowment, 2002.
23
6. J. Euzenat. Algebras of ontology alignment relations. In A. P. Sheth, S. Staab, M. Dean,
M. Paolucci, D. Maynard, T. W. Finin, and K. Thirunarayan, editors, International Seman-tic Web Conference, volume 5318 of Lecture Notes in Computer Science, pages 387–402.
Springer, 2008.
7. J. Euzenat, A. Ferrara, L. Hollink, A. Isaac, C. Joslyn, V. Malaise, C. Meilicke, A. Nikolov,
J. Pane, M. Sabou, F. Scharffe, P. Shvaiko, V. Spiliopoulos, H. Stuckenschmidt, O. Svab-
Zamazal, V. Svatek, C. T. dos Santos, G. A. Vouros, and S. Wang. Results of the ontology
alignment evaluation initiative 2009. In OM, 2009.
8. A. Ghazvinian, N. F. Noy, C. Jonquet, N. Shah, and M. A. Musen. What four million map-
pings can tell you about two hundred ontologies. In ISWC ’09: Proceedings of the 8thInternational Semantic Web Conference, pages 229–242. Springer-Verlag, 2009.
9. A. Ghazvinian, N. F. Noy, and M. A. Musen. Creating mappings for ontologies in
biomedicine: Simple methods work. In AMIA Annual Symposium (AMIA 2009), San Fran-
cisco, CA, 2009.
10. W. V. Hage, A. Isaac, and Z. Aleksovski. Sample evaluation of ontology-matching systems.
Proceedings of the ISWC workshop on Evaluation of Ontologies and Ontology-based tools,
pages 41–50, 2007.
11. W. Hu and Y. Qu. Falcon-AO: A practical ontology matching system. Web Semant.,6(3):237–239, 2008.
12. Y. R. Jean-Mary, E. P. Shironoshita, and M. R. Kabuka. ASMOV: Results for OAEI 2009.
In OM, 2009.
13. P. Lambrix, H. Tan, and Q. L. 0002. Sambo and sambodtf results for the ontology alignment
evaluation initiative 2008. In OM, 2008.
14. J. Madhavan and A. Y. Halevy. Composing mappings among data sources. In VLDB ’2003:Proceedings of the 29th international conference on Very large data bases, pages 572–583.
VLDB Endowment, 2003.
15. A. Miles and S. Bechhofer. Skos simple knowledge organization system reference, August
2009.
16. M. Nagy, M. Vargas-Vera, and P. Stolarski. Dssim results for oaei 2009. In OM, 2009.
17. N. F. Noy, N. H. Shah, P. L. Whetzel, B. Dai, M. Dorf, N. Griffith, C. Jonquet, D. L. Rubin,
M.-A. D. Storey, C. G. Chute, and M. A. Musen. Bioportal: ontologies and integrated data
resources at the click of a mouse. Nucleic Acids Research, 37(Web-Server-Issue):170–173,
2009.
18. A. Tordai, J. van Ossenbruggen, and G. Schreiber. Combining vocabulary alignment tech-
niques. In K-CAP ’09: Proceedings of the fifth international conference on Knowledge cap-ture, pages 25–32, New York, NY, USA, 2009. ACM.
19. A. Tordai, J. van Ossenbruggen, G. Schreiber, and B. Wielinga. Aligning large SKOS-like
vocabularies: Two case studies. In The Semantic Web: Research and Applications, volume
5554 of Lecture Notes in Computer Science, pages 198–212. Springer Berlin / Heidelberg,
2010.
20. P. Vossen, I. Maks, R. Segers, and H. van der Vliet. Integrating lexical units, synsets and
ontology in the Cornetto database. In E. L. R. A. (ELRA), editor, Proceedings of the SixthInternational Language Resources and Evaluation (LREC’08), 2008.
21. S. Wang, A. Isaac, B. A. C. Schopman, S. Schlobach, and L. van der Meij. Matching multi-
lingual subject vocabularies. In M. Agosti, J. L. Borbinha, S. Kapidakis, C. Papatheodorou,
and G. Tsakonas, editors, ECDL, volume 5714 of Lecture Notes in Computer Science, pages
125–137. Springer, 2009.
22. X. Zhang, Q. Zhong, F. Shi, J. Li, and J. Tang. Rimom results for oaei 2009. In OM, 2009.
24
Chinese Whispers and Connected Alignments
Oliver Kutz1, Immanuel Normann2, Till Mossakowski3, and Dirk Walther4
1 Research Center on Spatial Cognition (SFB/TR 8), University of Bremen, [email protected]
2 Department of Linguistics and Literature, University of Bremen, [email protected]
3 DFKI GmbH and SFB/TR 8 Spatial Cognition, University of Bremen, [email protected]
4 Faculty of Informatics, Technical University of Madrid (UPM), [email protected]
Abstract. This paper investigates the idea to treat repositories of ontologies as inter-linked networks of ontologies, formally captured by the notion of a hyperontology. Weapply standard matching algorithms to automatically create the linkage structure of therepository by performing pairwise matching. Subsequently, we define a modular work-flow to construct combinations of alignments for any finite number of ontologies. Thisworkflow employs and makes interoperable several tools from the ontology engineeringworld, comprising matching, reasoning, and structuring tools, and supports in particularmodular ontology extraction based on alignment, and a study and empirical analysis of(in)consistency propagation in connected alignments (the Chinese Whispers problem).
Keywords: Hyperontologies; Connected Alignments; Modularity; Consistency
1 Introduction and Problem Description
Ontology matching and alignment based on statistical methods is a relatively developedfield, with yearly competitions since 2004 comparing the various strengths and weaknessesof existing algorithms.5 In this paper, we aim at exploring the degrees to which statisticalalignment may lead to inconsistency in the merged ontologies. More precisely, we aim atinvestigating the effects, both theoretically and practically, of connected alignment,i.e. aligning several ontologies that match (non-trivially) pairwise. Our general approachis to treat large repositories of ontologies (in the order of hundreds of ontologies) as ourstarting point to perform pairwise matching in order to obtain an interlinked network ofontologies. Formally, such networks are captured by the notion of a hyperontology [6].Our work in progress is intended to answer questions such as the following:
Assuming pairwise alignments are consistent, how, and when, can we align fur-ther ontologies (in various orders) before we drift into inconsistency? In particular,how, and when, can we reduce the question of consistency of aligned ontologies andsatisfiability of matched concepts to the consistency of aligned sub-ontologies (i.e.modules generated by the matched sub-signatures)?
5 See http://oaei.ontologymatching.org/2009/
25
We here set up the theoretical background and necessary engineering environmentto give meaningful answers to such questions. In our related paper [9], we have studiedtechniques of information hiding to support the visualisation of the linkage structure andto allow a user to explore the complex networks resulting from pairwise matching onlarge sets of ontologies. We here focus on the last question mentioned above, namely howto reduce the consistency problem in an aligned network of ontologies to the consistencyof merged modules generated by respective alignments, and the corresponding interoper-ability problem between matching, modularity, and structuring tools. We also study theeffects of matching ontologies in different orders by looking at some specific examples.
Synonymy and Alignment as Colimit Computation. An essential part of thematching and alignment process is to relate and identify signature elements from differentontologies (possibly formulated in different ontology languages). Formally, this is capturedby the notion of a signature morphism.6 In the case of OWL ontologies, these aretype-preserving symbol mappings of the form σ : Sig(O1) → Sig(O2), i.e. mapping thesignature of O1 (= Sig(O1)) to that of O2, i.e. concepts to concepts, individuals toindividuals, and roles to roles.7 For a fixed ontology language, a signature morphismstraightforwardly induces a sentence translation map.
V-Alignments [13] abstractly capture the alignment process for synonymous signa-ture elements. Given ontologies O1 and O2, an interface (for O1, O2)
⟨Σ, σ1 : Σ−→Sig(O1), σ2 : Σ−→Sig(O2)
⟩
specifies that (using informal but suggestive notation)
– concepts σ1(c) in O1 and σ2(c) in O2 are identified for each concept c in Σ, regardlessof whether the concepts have the same name or not, and
– concepts in O1 \ σ(Σ1) and O2 \ σ(Σ2) are kept distinct, again regardless of whetherthey have the same name or not.
The resulting ontology O is not given a priori, but rather it is computed from the alignedontologies via the interface. This computation is a pushout in the sense of categorytheory, which in this case is just a disjoint union with identification of specific parts(namely those given through 〈Σ, σ1, σ2〉).
V-alignments can deal with basic alignment problems such as synonymy (identify-ing different symbols with the same meaning) and homonymy (separating (accidentally)identical symbols with different meaning)—see Fig. 1.
Example 1. In Fig. 1, the interface 〈Σ, σ1, σ2〉 specifies that the two instances of theconcept Woman as well as Person and Human are to be identified. This yields two conceptsWoman and Human_Being in the push-out ontology O obtained along the dashed arrows.It also determines that the two instances of Bank are to be understood as homonyms,and thus generates two new distinct concepts. �6 See e.g. [5] for the general institutional definition.7 We use the DL terminology concept name and role interchangeably with the OWL terminology class
and property.
26
Notion such as polysemy, however, are typically understood to relate terms that havea different, but related meaning, and can thus not be dealt with by simply identifyingsymbols or keeping them apart.8 Similarly, [13] raise the criticism that V-Alignments donot cover the case where a concept Woman in O1 is aligned with a concept Person in O2:here, the merged ontology should turn Woman into a subconcept of Person.
{Woman,River_Bank,Financial_Bank,Human_Being}
�
O
O1
�
O2
�
�{Woman,Bank,Person} �{Woman,Bank,Human}Σ
σ2
��σ1
={Woman_Woman,Person_Human}
Fig. 1. V-alignment: merge through interface (dashed arrows are automatically computed via colimits)
Whilst this is not directly possible with pushouts, we are here only interested inmatching synonyms across a network of ontologies, and for this purpose, V-alignments(and their compositions) are sufficient. Studying more complex alignment operations weleave for future work. We next turn to the problem of aligning several ontologies at once.
Consistency and Chinese Whispers. The game of Chinese Whispers9 is playedas follows: n persons are arranged in a certain (typically circular) order such that for eachperson Pi there is a j such that Pi exchanges a message with Pj . The point of the gameis to observe the distortion of the message as it travels from P1 along the communicationchannel. We here are interested in the effects of playing Chinese Whispers with ontologies,where the pairwise matching replaces the transmission of a message, i.e. the messagesbeing exchanged are of the form: “Oi and Oj agree that concept C of Oi is synonymouswith concept D of Oj”.
We make the following idealisations concerning ‘matching’ a) we assume that inpairwise matching the order does not matter, i.e. matching O1 with O2 yields the samecolimit ontology (i.e. alignment) as matching O2 with O1
10; b) matching algorithms are‘not transitive’, i.e., matching 〈O1, O2〉 and 〈O2, O3〉 and computing the colimit yields,in general, a different result than matching and aligning 〈O1, O3〉,11 c) we assume thatwe do not match ontologies with themselves.12
8 This problem can be addressed by considering E-connections as a general form of alignment (see [6]).9 In the United States, “Telephone” is the most common name for the game. The name “Chinese whis-
pers” reflects the former stereotype in Europe of the Chinese language as being incomprehensible.Although it is sometimes considered offensive in the US, it remains the common British English namefor the game and is not generally regarded as being offensive.
10 Whether or not this holds for actual matching systems is an implementational artefact which we ignore;the assumption is certainly reasonable to make as both ‘agreement’ and ‘synonymy’ are symmetric.
11 In other words, the composition of the two alignments (the composition operation is easily seen tobe associative via a pullback operation, see [13]) will typically not agree with a matcher’s resultscomparing O1 and O3 directly.
12 Although one would expect to get the identity matching in such a case, actual matching tools behavesometimes rather unpredictable in these cases and often return no results.
27
With these assumptions in place, given a repository R with N ontologies, we startwith l = N×(N−1)
2 matching pairs.Playing chinese whispers on R with k ≤ N players now means to pick a connected
subgraph of the hyperontology graph (to ensure that each ontology ‘talks’ to at least oneother), which we call a matching configuration . Note that, by assumption, matchingconfigurations contain no loops (i.e. reflexive vertices), are undirected (because of theassumed symmetry of the matching results), and contain at most one edge between twovertices. Therefore, matching configurations are connected simple graphs.
k = 3
k = 2
Fig. 2. The number of non-isomorphic matching configurations for k = 2, 3
Given a fixed k, the largest possible matching configuration (measured in pairs ofmatched ontologies, i.e. edges) corresponds to a clique with k nodes, i.e. a completesubgraph of the hyperontology graph with k nodes: such a graph has
∑k−1i=1 i edges.
In practise, not every pair of ontologies will match, i.e. a matcher will report no syn-onyms. Therefore, for fixed k ≤ N ontologies O1, . . . , Ok, the number of non-isomorphicmatching configurations containing the Oi (i = 1, . . . , k) corresponds to the number ofconnected components of the hyperontology graph with these ontology nodes. The casesof N = 2, 3 (assuming a clique) are illustrated in Fig. 2. The alignment operation on amatching configuration, i.e. the computation of the colimit of that graph, we call a con-nected alignment. The Chinese Whispers Problem is now the following question:
For what kinds (or shapes) of matching configurations and exchanged matchingresults does the consistency of the input ontologies propagate to the merge (colimit)of the matching configuration?
We will give some partial answers to this problem in Sec. 2 by showing that theconsistency of an aligned matching configuration is reducible to the consistency of thealignment of ‘reasonably’ large modules talking only about the matched signatures. InSec. 3, we will discuss in detail an alignment of three ontologies involving the Dolceontology, with pairwise consistent alignments, but an overall inconsistent one. Moreover,in Sec. 4, we will describe how the results of a matcher comparing ontologies O1 and O2
(and giving rise to a V-alignment) can be rewritten into a structured ontology for furtherprocessing with our tool Hets introduced below, and describe the workflow employingstandard ontology matching and reasoning tools. Finally, Sec. 5 describes related andfuture work.
28
2 Modularity in Hyperontologies
In the following, we will make precise what we mean by a module and define the notionof conservativity. We start with some auxiliary notions. Let Σ be a signature containingconcept names and roles. Let Sen(Σ) be the set of sentences formulated using the symbolsin Σ in some language. An ontology O in signature Σ is then simply a subset of Sen(Σ).The sentences of course depend on the language the ontologies under consideration areformulated in, e.g. OWL or some fragment thereof.
We continue with introducing a general notion of a module in the sense that a moduleof an ontology is not restricted to be a subset of the ontology. It is crucial, however, thatthe module says everything (expressible in its signature) that is said by the ontology itself(i.e. the ontology is required to be a conservative extension of the module)—see Fig. 3 .
subset of axioms translation along signature morphism
σ : T1 −→ T2
T
T1
T2
Fig. 3. Modules as subsets vs. modules as image under translation.
Definition 1. A theory morphism σ : O1 −→ O2 is consequence-theoretically con-servative, if O2 does not entail anything new w.r.t. O1, formally, O2 |= σ(ϕ) impliesO1 |= ϕ. Moreover, σ : O1−→O2 is model-theoretically conservative, if any O1-modelM1 has a σ-expansion to O2, i.e. a O2-model M2 with M2|σ = M1.
Here, |= as usual denotes logical consequence, whereas _|σ denotes model reduct fora signature morphism σ : Σ1 → Σ2, i.e. for a Σ2-model M2, M2|σ is a Σ1-model thatinterprets a symbol by first translating it along σ and then interpreting it using M2.
It is easy to show that conservative theory morphisms compose. Moreover, the notionof model-theoretic conservativity is stronger than consequence-theoretic conservativity.To be precise, the former implies the latter, but not vice versa [7]. The two notionscoincide if we define consequence-theoretic conservativity using Σ-theories that containconsequences φ ∈ Sen(Σ) formulated in second-order logic.
The computational complexity of deciding conservativity appears to be rather daunt-ing even if the ontologies are formulated in weak logics. For instance, for ontologiesformulated in the light-weight Description Logic EL, deciding consequence-theoretic con-servativity is ExpTime-complete, and model-theoretic conservativity is undecidable. Theformer problem also becomes undecidable when adding nominals to ALCIQ, for which
29
it is still 2-ExpTime-complete [8]. This suggests that, for practical purposes and appli-cations, we often have to live with approximations of these notions, more precisely withsufficient (syntactic) conditions for conservativity that allow to construct non-minimalmodules. Indeed, the notion of an ontology module of an ontology O has been defined asany “subontology O′ such that O is a conservative extension of O′” [1].
Definition 2 (Module Generator). Let O be an ontology in some fixed DL, and letΣ ⊆ Sig(O) be a signature. Sen(Sig(O)) is the set of sentences in Sig(O). A function
Π : 〈O,Σ〉 �→ Sen(Sig(O))
mapping pairs 〈O,Σ〉 consisting of an ontology O together with a signature Σ to a set ofsentences in Sig(O) is called a Σ-module generator if for all O and Σ:
Π(〈O,Σ〉) is a model-theoretic Sig(O)-conservative extension of O.
For a Σ-module generator Π, the set Π(〈O,Σ〉) is called a Σ-module for O.Π(〈O,Σ〉) is called Σ-covering for O if:
O is a model-theoretic Σ-conservative extension of Π(〈O,Σ〉).
The idea to use (conservative) module generators is to massively reduce the sizeof a colimit ontology to a merge of modules generated by the matched signatures andpreserving the semantics completely. This means that we can check the satisfiability ofour matched concepts (and the consistency of the overall merged ontology) already in arather small fragment of the overall ontology. Indeed, it is not hard to construct (or find)cases where ontologies have a moderate semantic overlap, but where the overall mergewill be very hard to process by current tools.
Before coming to our central theoretical result, we need some preparatory notions.
Definition 3. A diagram is a graph D of signatures (Di)i∈D and signature morphisms(Dm : Di → Dj)m : i→j∈D. Given a diagram D, a family of models Mi ∈ Mod(Di)i∈Dis compatible, if Mj |Dm = Mi for any m : i → j ∈ D. A logic has the amalgamationproperty w.r.t. a class of diagrams if for any diagram in the class, any compatible familyof models can be amalgamated to a unique model of the colimit of the diagram (i.e. suchthat the reducts along the colimit injections yields the models of the family).
A logic is semi-exact, if it has the amalgamation property for pushout diagrams.
Note that colimits of theories can be easily defined in terms of signature colimitsand unions of (translated) axioms; the amalgamation property then carries over fromsignature colimits to theory colimits (see [11], 4.4.17). We formulate the next results intheir full generality to cover also ontology languages other than DLs. But note that theyapply in particular to all usual DLs as these are semi-exact and moreover have an initialsignature (i.e. a signature with a unique signature morphism into any signature) [5].
Theorem 1 (Combination of two modules). Assume a semi-exact logic. ConsiderFig. 4, and assume that O and O are obtained by pushouts (with base Σ). Then O is amodel-theoretic conservative extension of O. In particular, O is satisfiable iff O is.
30
O
O1
�
O2
�
˜Oc?
�
Π(O1, Σ1)
c�
�
Π(O2, Σ2)
c�
�
Σ
σ2��
σ1
Fig. 4. Propagation of modular structure through one matching (a ‘c’ denotes conservativity)
Proof. Let M be a model of O. For i = 1, 2, let Mi be an Oi-expansion of the Π(O1, Σi)-reduct of M (which exists by conservativity of Oi over Π(O1, Σi)). M1 and M2 obvi-ously agree on Σ, hence they have an amalgamation M which is an O-model. Now theΠ(O1, Σi)-reducts of M agree with those of M , hence by uniqueness of amalgamation,the O-reduct of M is M . �
We next generalise this result to the case of arbitrary matching configurations.
Theorem 2 (Combination of multiple modules). Assume a semi-exact logic withan initial signature. Consider a family of ontologies (Oi)i∈I indexed by a finite non-emptyset I and a simple graph G ⊆ I × I, such that for (i, j) ∈ G, Oi and Oj are interfaced by
Oi� θi,j
Σi,jθj,i � Oj
DefineΣi :=
⋃j∈I\{i}
θi,j(Σi,j) (1)
and σi : Σi → Π(Oi, Σi) the module in Oi for Σi. Let σi,j : Σi,j → Π(Oi, Σi) be therestriction of θi,j, namely θi,j : Σi,j → Σi composed with σi.13 Assume that O (resp.O) is obtained by the colimit of the diagram of all σi,j (resp. all σi,j composed with theinclusion of Π(Oi, Σi) into Oi). Then O is a model-theoretic conservative extension ofO. In particular, O is satisfiable iff O is.
Proof. Note that the diagrams for obtaining O resp. O can be turned into connecteddiagrams by adding the inclusions of the empty signature into all involved theories (thecolimit does not change by this addition). The rationale is that this ensures in OWLthat all compatible model families are built over the same universe of individuals. Moreformally, by Prop. 4.4.15 of [11], in any semi-exact logic with an initial signature, allfinite non-empty connected diagrams enjoy the amalgamation property. With this, theproof is a straightforward generalisation of the proof of Thm. 1. �13 Consider Fig. 4, but replace Σ by Σi,j , σ1 by σi,j , σ2 by σj,i, O1, Σ1 by Oi, Σi, and O2, Σ2 by Oj , Σj .
31
Note that Thm. 1 does not hold for consequence-theoretic conservativity. Considerthe following example, adapted from [7]. In ALCO, let O1 be
IntroTCS � ∃has_subject.AutomataTheoryIntroTCS � ∃has_subject.ComplexityTheoryAutomataTheory � ComplexityTheory � ⊥
and O2 beIntroTCS � ∀has_subject.{moore_automata}IntroTCS � ∃has_subject.{moore_automata}
Let Σ be {IntroTCS, has_subject}. Then assuming a consequence-theoretically conser-vative (minimal) module generator, Π(O1, Σ) = Π(O2, Σ) = O is
IntroTCS � ∃has_subject.
But this is consistent, while O1 ∪O2 is not (assuming IntroTCS is also instantiated).
3 Worked Out Example
The following worked out example serves two purposes: it should illustrate the last theo-rem on combination of multiple modules, and it should demonstrate its practical impact.
Our ontology repository ORATE14 contains among others three ontologies, namely:SpatialRelations, ExtendedDnS, and FunctionalParticipation. Let us denote themby O1, O2, and O3 resp., to be notationally conform with Theorem 2 above. We combinethem in different ways and show how consistency issues of the combined ontologies canalready be answered in the combination of modules. Automated matching resulted in theconcept identifications listed in the two tables below.
SpatialRelations endurant participant place-ofExtendedDnS physical-endurant participant-place-of situation-place-of
Table 1. Matching SpatialRelations against ExtendedDnS
Our first matching thus induces an interface 〈Σ12, θ12, θ21〉 with
Σ12 = {endurant, participant, place-of},
θ12 = id, and θ21 = {endurant �→ physical-endurant, participant �→ paticipant-place-of,place-of �→ situation-place-of}.
The second an interface 〈Σ23, θ23, θ32〉 with Σ23 = {endurant, physical-object, region},θ23 = {endurant �→ non-physical-endurant, physical-object �→ agentive-physical-object,region �→ space-region}, and θ32 = id.
From Σ12 and Σ23, the Σi’s can be determined: Σ1 = θ12Σ12 = {physical-endurant},Σ2 = θ12Σ12 ∪ θ23Σ23 = {endurant}, and Σ2 = θ23Σ32 = {non-physical-endurant}.14 http://ontologies.informatik.uni-bremen.de
32
SpatialRelations endurant physical-object regionFunctionalParticipation non-physical-endurant agentive-physical-object space-region
Table 2. Matching SpatialRelations against FunctionalParticipation
The signatures Σi, for i = 1, 2, 3, together with corresponding ontologies Oi are sentto the module generator that gives us for each ontology the corresponding moduleMi := Π(Oi, Σi). Finally, from the signatures Σi and the signature morphisms θij , thecolimit (i.e. the alignment) M of the modules M1, M2, and M3 is obtained. Similarly,the aligned ontology M can be determined from the three ontologies O1, O2, and O3.A practical result of this alignment can be a check of the merged module M for consis-tency. As we would expect, the merge of the concept "physical-endurant", "endurant",and "non-physical-endurant" into a single concept makes it unsatisfiable—this result canbe automatically verified by a prover. Since we know that O must be a conservativeextension of M , it is in fact not necessary to build O and check its consistency: thisinformation can already be inferred from M . To repair the detected inconsistency of M ,we can abandon either M1 or M3 from the alignment process. In both cases, M1 alignedwith M2 only, or M3 aligned with M2 only, the resulting merged module turns out to beconsistent, and we know by conservativity that the corresponding ontologies would beconsistent, too.
4 An Interoperability Workflow and Prototypical Implementation
4.1 The Component Tools
We implemented a workflow for aligning arbitrary matching configurations taking advan-tage of several third-party tools. We here briefly introduce these tools and describe in thesubsequent subsection their interoperation. Our ontologies to be matched and aligned aretaken from our ontology repository ORATE which is being developed and maintainedwithin the EU-project OASIS15. The software is based on BioPortal [10]. As matchingsystem we use Falcon [3] which matches OWL ontologies by means of linguistic andstructural analysis. Falcon can be comfortably used in a batch mode and thus makes itsuitable for a pipe workflow. For module extraction as well as consistency checks we usePellet [12] which in particular makes use of the OWL-API 16. Finally, we use Hets 17 forthe computation of colimits (i.e. alignments). Hets is a parsing, static analysis and proofmanagement tool incorporating various provers and different specification languages, thusproviding a tool for heterogeneous specifications.
4.2 Workflow Description
Our workflow of multiple ontology alignment consists of two phases: 1) the preprocessingof the whole repository (ORATE) to a complete list of pairwise matching records15 See http://www.oasis-project.eu16 See http://owlapi.sourceforge.net17 See www.informatik.uni-bremen.de/cofi/hets
33
and 2) the alignment of a connected subgraph of the hyperontology graph. It is upto user what part of the hyperontology graph the user selects as connected subgraph.The rationale behind this user interaction is to dismiss false matchings produced bythe matching system. Different matching configuration consequently lead to differentalignment outcomes.
preprocess_repository = {foreach (o1,o2) in repositorymatch_ontologies o1 o2
end}
match_ontologies o1 o2 = {mapping = falcon o1 o2cs1 = extract_concepts_from_o1 mappingcs2 = extract_concepts_from_o2 mappings12 = build_interface_signature cs1 cs2sm1 = build_interface_signature_morphism s12 cs1sm2 = build_interface_signature_morphism s12 cs2matching_record = (o1,o2,s12,sm1,sm2)if (not_empty mapping) store_in_hyperontology_graph matching_record
}
compute_module_graph hyperontology_subgraph = {foreach node in hyperontology_subgraph
in_edges = edges incident with nodesig = compute_interface_signature node in_edgesmodule = compute_module node sig
endreplace nodes in hyperontology_subgraph by modules
}
align_modules hyperontology_subgraph = {module_graph = compute_module_graph hyperontology_subgraphinterfaces = {compute_interface edge | edge in module_graph}views = {compute_views edge | edge in module_graph}spec = write_alignment_spec modules interfaces viewsalignment = compute_alignment spec % hets
}
Fig. 5. Pseudo code of the workflow for multiple ontology alignment
Fig. 5 shows the whole workflow in pseudo code. We are going to explain it now lineby line. Procedure preprocess_repository takes each pair of ontologies (o1,o2) andapplies the procedure match_ontologies to it, i.e., it matches pairwise all ontologies fromthe repository. The output of match_ontologies is a matching_record: the matchingsystem Falcon computes the mapping between the two ontologies o1 and o2. From themapping the concepts cs1 (cs2) belonging to ontology o1 (o2) are extracted. Based on theconcepts cs1 (cs2), the interface signature s12 is built and the corresponding signaturemorphisms sm1 and sm2 to the ontologies o1 and o2. The two ontologies, the interfacesignature, and the two signature morphisms form the matching_record. This recordcan be viewed as a link between two ontologies. We call this network whose nodes are
34
ontologies and whose edges are the matching records hyperontology graph. Although allontologies are matched pairwise, the graph is not complete, i.e., some pairs of ontologies(in practice even the majority) are not linked, namely when the matching system cannotfind any mappings.
Once the hyperontology graph of the ontology repository is computed we can choosean arbitrary subgraph of it to compute the colimit of modules implicitly given in this sub-graph. For that the procedure compute_module_graph takes the hyperontology subgraphand basically replaces its ontologies by modules extracted from them. More precisely, foreach ontology (=node) it collects all interfaces (=in_edges) connected to this node andcomputes (according to Equation 1 in Thm. 2) a signature (sig) that is finally used tocompute the module. Practically, this last step is delegated to Pellet (OWL-API).
Aligning modules implicitly given in a hyperontology subgraph (cf. align_modulesprocedure) comprises the following steps: we first transform the graph with the justmentioned procedure compute_module_graph to the module_graph. From the modulegraph we extract all signatures (=interfaces) and signature morphisms (views) to themodules. From the modules, the interfaces, and the views we compose a specificationdocument (spec) that can be understood by Hets. Finally, Hets computes the colimit(alignment) of this spec.
5 Related Work and an Outlook
Matching and revision in networks of ontologies seems to be a rather unexplored topic.Although only studying pairs of ontologies, the most closely related work in spirit appearsto be [4], where a semi-automatic procedure is presented for the integration of ontologiesthat involves revision of mappings. This approach is implemented as the Protégé 4 pluginContentMap. Here, a selected ontology matcher is used to compute mappings betweenthe signatures of two ontologies chosen for integration, the mappings are explicitly inter-nalised as axioms in an OWL-2 ontology, and the result of the integration is then takento be the (disjoint) union of the original ontologies together with the mapping axioms.The integration is assumed to be successful if a user does not identify unintended logicalconsequences. This decision is guided by justifications (explanations for entailment) thatcan automatically be computed, e.g., by ContentMap and the confidence values createdby the matcher for the mapping axioms. To eliminate the unintended consequences, a re-pair plan is created describing which axioms from the original ontologies or the mappingshould be removed. It should be noted that such a plan does not always exist and thata desired integration may require the iteration of these steps.
The main differences to our approach are a) we support heterogeneity of ontologylanguages, b) we do not internalise the mappings but make explicit the structure ofthe alignment graph, c) we generically support arbitrary matching configurations anduse category-theoretic techniques to compute the merged ontologies without duplicatingmatched signature items. Apart from these differences, the necessity for debugging andrevising matchings also applies to our approach, and the proposed techniques can bemade fruitful also in this setting.
35
We could here only scratch the surface of the area of problems related to matchingin networks of ontologies. We have laid out the necessary engineering infrastructure tocombine matching, structuring, and reasoning tools, and obtained some theoretical resultsconcerning the reduction of consistency checks to merged modules.
However, a lot of open questions remain. For instance, internalising mappings can beextended to internalising the confidence values of mappings building on similarity-basedE-connections [2]. Moreover, a full statistical analysis of major ontologies and repositoriesremains to be done to understand the impact of iterated matching on consistency.
Acknowledgements
Work on this paper has been supported by the DFG-funded collaborative research centreSFB/TR 8 ‘Spatial Cognition’, the EU-funded OASIS project, and by the German FederalMinistry of Education and Research (Project 01 IW 07002 FormalSafe). Dirk Walther issupported by a ‘Juan de la Cierva’ postdoctoral fellowship.
References
1. Ghilardi, S., Lutz, C., and Wolter, F. Did I Damage My Ontology? A Case for ConservativeExtensions in Description Logics. In Proceedings of KR-06 (2006), pp. 187–197.
2. Hois, J., and Kutz, O. Counterparts in Language and Space—Similarity and S-Connection. InProc. of FOIS 2008 (2008), C. Eschenbach and M. Grüninger, Eds., IOS Press, pp. 266–279.
3. Hu, W., and Qu, Y. Falcon-AO: A practical ontology matching system. In Proc. of WWW-07(2008), pp. 237–239.
4. Jimenez Ruiz, E., Cuenca Grau, B., Horrocks, I., and Berlanga, R. Ontology IntegrationUsing Mappings: Towards Getting the Right Logical Consequences. In Proc. of the 6th EuropeanSemantic Web Conference (ESWC 2009) (2009), LNCS, Springer.
5. Kutz, O., and Mossakowski, T. Conservativity in Structured Ontologies. In 18th EuropeanConf. on Artificial Intelligence (ECAI-08) (Patras, Greece, 2008), IOS Press.
6. Kutz, O., Mossakowski, T., and Lücke, D. Carnap, Goguen, and the Hyperontologies: LogicalPluralism and Heterogeneous Structuring in Ontology Design. Logica Universalis 4, 2 (2010). Specialissue on ‘Is Logic Universal?’.
7. Lutz, C., Walther, D., and Wolter, F. Conservative Extensions in Expressive DescriptionLogics. In Proc. of IJCAI-07 (2007), M. Veloso, Ed., AAAI Press, pp. 453–458.
8. Lutz, C., and Wolter, F. Conservative Extensions in the Lightweight Description Logic EL. InProc. of CADE-07 (2007), Springer, pp. 84–99.
9. Normann, I., and Kutz, O. Ontology Reuse and Exploration via Interactive Graph Manipulation.In Proc. of the ISWC Workshop on Ontology Repositories for the Web (SERES-2010) (ISWC-2010,November 7, 2010, Shanghai, China), CEUR-WS.
10. Noy, N. F., Shah, N. H., Dai, B., Dorf, M., Griffith, N., Jonquet, C., Montegut, M. J.,
Rubin, D. L., Youn, C., and Musen, M. A. Bioportal: A web repository for biomedical ontologiesand data resources [demonstration], 2007.
11. Sannella, D., and Tarlecki, A. Foundations of Algebraic Specifications and Formal ProgramDevelopment. Springer Verlag. to appear.
12. Sirin, E., Parsia, B., Cuenca Grau, B., Kalyanpur, A., and Katz, Y. Pellet: A practicalOWL DL reasoner. Journal of Web Semantics 5, 2 (2007), 51–53.
13. Zimmermann, A., Krötzsch, M., Euzenat, J., and Hitzler, P. Formalizing Ontology Align-ment and its Operations with Category Theory. In Proc. of FOIS-06 (2006), pp. 277–288.
36
Consistency-driven Argumentationfor Alignment Agreement
Cassia Trojahn and Jerome Euzenat
INRIA & LIG, 655 Avenue de l’Europe, Montbonnot Saint Martin, France{cassia.trojahn,jerome.euzenat}@inrialpes.fr
Abstract. Ontology alignment agreement aims at overcoming the problem thatarises when different parties need to conciliate their conflicting views on ontol-ogy alignments. Argumentation has been applied as a way for supporting the cre-ation and exchange of arguments, followed by the reasoning on their acceptabil-ity. Here we use arguments as positions that support or reject correspondences.Applying only argumentation to select correspondences may lead to alignmentswhich relates ontologies in an inconsistent way. In order to address this problem,we define maximal consistent sub-consolidations which generate consistent andargumentation-grounded alignments. We propose a strategy for computing theminvolving both argumentation and logical inconsistency detection. It removes cor-respondences that introduce inconsistencies into the resulting alignment and al-lows for maintaining the consistency within an argumentation system. We presentexperiments comparing the different approaches. The (partial) experiments sug-gest that applying consistency checking and argumentation independently sig-nificantly improves results, while using them together does not bring so much.The features of consistency checking and argumentation leading to this result areanalysed.
1 Introduction
Due to the diverse ways of exploring the ontology matching problem, matching systemsgenerally differ in the alignments generated between two ontologies. Some approachesmay be better suited for some ontologies, or some tasks, than others. Ontology align-ment agreement aims at overcoming the problem of allowing different parties to con-ciliate their conflicting points of view on alignments. There may be different ways toperform alignment agreement, such as voting or weighting. In this paper, we considerargumentation which offers a more reasoned way to decide which correspondences topreserve.
Argumentation theory has been exploited as a way to support the comparison andselection of correspondences within an alignment process. Correspondences are rep-resented as arguments and argumentation frameworks support the reasoning on theiracceptability. This approach has been used in different scenarios. [13] propose an ap-proach for supporting the creation and exchange of different arguments, that support orreject correspondences in the context of agent communication. In [18], different match-ers work on the basis of particular approaches achieving distinct results that are com-pared and agreed via an argumentation process.
37
An open issue in alignment agreement is related to the inconsistency in the agreedalignment. Indeed, some selected sets of correspondences may relate the ontology in aninconsistent way. Most matching systems do not consider logic-based semantics in theiralgorithms. As a result, almost all matching systems produce incoherent alignments[14]. Although argumentation aims at resolving conflicts on the alignments generatedby these systems, this process does not guarantee that the agreed alignment is consistenteven if the initial alignments were consistent.
In this paper, we propose a model that involves both argumentation and logical in-consistency detection. We focus on the scenario where matchers working on the basisof different matching approaches try to reach a consensus on their alignments. First,matchers generate their correspondences, representing them as arguments. Next, theyexchange their arguments and interpret them under argumentation frameworks based ontheir individual preferences. The arguments in every individual set of acceptable argu-ments are considered as an agreed alignment. Then, the inconsistent correspondencesin such sets are removed, in order to generate a maximal consistent agreed alignment.This allows for maintaining the consistency within an argumentation system. We eval-uate our proposal on a standard set of alignments. Though theoretically grounded, theconsistency step does not improve argumentation alone. For some test cases, the ar-gumentation process is incidentally able to provide consistent agreed alignments. Wedescribe the features of consistency checking and argumentation which cause this re-sult.
The rest of the paper is organised as follows. First, we introduce alignments and in-consistency of alignments (§2). We then present the argumentation approach for align-ment agreement (§3). Next, the consistency-driven argumentation protocol is presented(§4) and its evaluation is discussed (§5). Finally, we discuss related work (§6) and con-clude the paper (§7).
2 Alignments and Inconsistency
An alignment (A) is a set of correspondences from a pair of ontologies (o and o′). Eachcorrespondence is a quadruple: 〈e, e′, r, n〉, where e ∈ o, e′ ∈ o′, r is the relation be-tween e and e′, taken from set of alignment relations (e.g., ≡, �, � or ⊥), and n ∈ [0 1]is a confidence level (e.g., measure of confidence in the fact that the correspondenceholds). For instance, given the two ontologies of Figure 1, one can consider the follow-ing correspondences, meaning that (1) the two classes Person in both ontologies are thesame, and that (2) DepartmentHead in the first ontology is subsumed by Department inthe second ontology.
〈Persono, Persono′ ,≡, 1.0〉(1)
〈DepartmentHeado, Departmento′ ,�, 0.8〉(2)
The semantics of alignments provides a definition of how alignments must be inter-preted. It is related to the semantics of the aligned ontologies, which is given by theirsets of models M(o) and M(o′). The main effect of alignments is to select compatiblepairs of models of the two related ontologies [8].
38
Person
hasAddressAddress
DepartmentHead
manages
Dpt
ProductLine
Person
hasContactContact
Employee
manages
Department
Product
⊗
=
�
Fig. 1. Fragments of ontologies o and o′ with alignment A.
We rely here on a basic semantics in which models are directly compatible. It con-siders that a correspondence is satisfied by a pair of models if the interpretation of theentities by these models satisfy the relation of the correspondence.
Definition 1 (Satisfied correspondence). A correspondence c = 〈e, e′, r〉 is satisfiedby two models m, m′ of o, o′ on a common domain D if and only if m ∈ M(o),m′ ∈ M(o′) and
〈m(e), m′(e′)〉 ∈ rU
such that rU ⊆ D×D is the interpretation of the relation. This is denoted as m, m′ |= c.
For instance, in the language used as example, if m and m′ are respective modelsof o and o′:
m, m′ |= 〈c, c′,≡〉 iff m(c) = m′(c′)
m, m′ |= 〈c, c′,�〉 iff m(c) ⊆ m′(c′)
m, m′ |= 〈c, c′,�〉 iff m(c) ⊇ m′(c′)
m, m′ |= 〈c, c′,⊥〉 iff m(c) ∩ m′(c′) = ∅
Definition 2 (Models of aligned ontologies). Given two ontologies o and o′ and analignment A between these ontologies, a model of these aligned ontologies is a pair〈m, m′〉 ∈ M(o)×M(o′), such that each correspondence of A is satisfied by 〈m, m′〉.
The alignment acts as a model filter for the ontologies: it selects the interpretation(here the models) of ontologies which are coherent with the alignments. This allows fortransferring information from one ontology to another since reducing the set of modelswill entail more consequences in each aligned ontology.
The notion of models of aligned ontologies is also useful for defining the usualnotions of consistency or consequence.
Definition 3 (Consistent alignment). Given two ontologies o and o′ and an alignmentA between these ontologies, A is consistent if there exists a model of A. Otherwise A isinconsistent.
39
For instance, under the classical ontology interpretation, the alignment A presentedin Figure 1 is inconsistent as soon as there exists a DepartmentHead because any modelwould require to satisfy the following equations:
m(Persono) = m′(Persono′) A
m(DepartmentHeado) ⊆ m′(Departmento′) A
m(DepartmentHeado) ⊆ m(Persono) o
m′(Departmento′) ∩ m′(Persono′) = ∅ o′
and the DepartmentHead would need to be in both the interpretation of Departmento′
and in that of Persono′ .In this paper we will only consider inconsistency, however, the same applies to
incoherence: the fact that a class or relation may necessarily be empty, i.e., which wouldcause inconsistency if instantiated.
3 Argumentation Approach
In alignment agreement, arguments can be seen as positions that support or reject corre-spondences. Such arguments interact following the notion of attack and are selected ac-cording to the notion of acceptability. These notions were introduced by [6]. In Dung’smodel, the acceptability of an argument is based on a reasonable view: an argumentshould be accepted only if every attack on it is attacked by an accepted argument. Dungdefines an argumentation framework as follows.
Definition 4 (Argumentation framework [6]). An Argumentation Framework (AF) isa pair 〈A, �〉, such that A is a set of arguments and � (attacks) is a binary relationon A. a � b means that the argument a attacks the argument b. A set of arguments S
attacks an argument b iff b is attacked by an argument in S.
In Dung’s model, all arguments have equal strength, and an attack always suc-ceeds (or successfully attacks). [2] has introduced the notion of preference betweenarguments, where an argument can defend itself against weaker arguments. This modeldefines a global preference between arguments. In order to relate preferences to differ-ent audiences, [3] proposes to associate arguments to the values which supports them.Different audiences can have different preferences over these values. This leads to thenotion of successful attacks, i.e., those which defeat the attacked argument, with respectto an ordering on the preferences that are associated with the arguments. This allowsfor accommodating different audiences with different interests and preferences.
Bench-Capon’s framework acknowledges the importance of preferences when con-sidering arguments. However, in the specific context of ontology matching, an objectioncan still be raised about the lack of complete mechanisms for handling persuasiveness[10]. Indeed, many matchers output correspondences with a strength that reflects theconfidence they have in the fact that the correspondence between the two entities holds.These confidence levels are usually derived from similarity assessments made duringthe matching process. They are therefore often based on objective grounds.
40
For associating an argument to a strength, which represents the confidence that anagent has in some correspondence, [18] has proposed the strength-based argumentationframework, extending Bench-Capon’s model:
Definition 5 (Strength-based argumentation framework (SVAF) [18]). A SVAF isa sextuple 〈A, �,V, v, , s〉 such that 〈A, �〉 is an AF, V is a nonempty set of values,v : A → V , is the preference relation over V (v1 v2 means that, in this framework,v1 is preferred over v2), and s : A → [0, 1] represents the strength of the argument.
Each audience α is associated with its own argumentation framework in which onlythe preference relation α differs. In order to accommodate the notion of strength, thenotion of successful attack is extended:
Definition 6 (Successful attack [18]). An argument a ∈ A successfully attacks (ordefeats, noted a†αb) an argument b ∈ A for an audience α iff
a � b ∧ (s(a) > s(b) ∨ (s(a) = s(b) ∧ v(a) α v(b)))
Definition 7 (Acceptable argument [3]). An argument a ∈ A is acceptable to anaudience α with respect to a set of arguments S, noted acceptableα(a, S), iff ∀x ∈ A,x†αa ⇒ ∃y ∈ S; y†αx.
In argumentation, a preferred extension represents a consistent position within aframework, which defends itself against all attacks and cannot be extended withoutraising conflicts:
Definition 8 (Preferred extension). A set S of arguments is conflict-free for an audi-ence α iff ∀a, b ∈ S,¬(a � b) ∨ a†αb. A conflict-free set of arguments S is admissiblefor an audience α iff ∀a ∈ S, acceptableα(a, S). A set of arguments S in the VAF is apreferred extension for an audience α iff it is a maximal admissible set (with respect toset inclusion) for α.
In order to determine preferred extensions with respect to a value ordering promotedby distinct audiences, objective and subjective acceptance are defined [3]. An argumentis subjectively acceptable if and only if it appears in some preferred extension for somespecific audience. An argument is objectively acceptable if and only if it appears inall preferred extension for every specific audience. We will call objective consolidationthe intersection of objectively acceptable arguments for all audiences and subjectiveconsolidation the union of subjectively acceptable arguments for all audiences.
3.1 Arguments on correspondences
A way of representing correspondences as arguments within an AF is as follows:
Definition 9 (Argument [13, 17]). An argument a ∈ A is a triple a = 〈c, v, h〉, suchthat c is a correspondence,v ∈ V is the value of the argument and h is one of +,-depending on whether the argument is that c does or does not hold.
The notion of attack is then defined as follow:
41
Definition 10 (Attack [13, 17]). An argument 〈c, v, h〉 ∈ A attacks an argument〈c′, v′, h′〉 ∈ A iff c = c′ and h �= h′.
For instance, if a = 〈c, v1, +〉 and b = 〈c, v2,−〉, a � b and vice-versa (b is thecounter-argument of a, and a is the counter-argument of b).
The way arguments are generated differs in each scenario. The strategy in [18],negative arguments as failure, relies on the assumption that matchers return completeresults. Each possible pair of ontology entities which is not returned by the matcher isconsidered to be at risk, and a negative argument is generated (h=-).
In this paper, different matchers argue with each others in order to obtain an agree-ment on their alignments. To do this, each matcher is a different audience. The valuesin V correspond to the different matching approaches and each matcher m has a prefer-ence ordering m over V such that its preferred values are those it associates to its ar-guments. For instance, consider V = {l, s, w}, i.e., lexical, structural and wordnet-basedapproaches, respectively, and three matchers ml, ms and mw, using such approaches.The matcher ml has as preference order l ml
s mlw.
To illustrate the agreement process, consider the alignment A of Figure 1 and twomatchers i and j. Both i and j generate the correspondence (1) and j the correspondence(2). The following arguments are then created by i and j:
ai,1 : 〈〈Persono, Persono′ ,≡, 1.0〉, w,+〉
ai,2 : 〈〈DepartmentHeado, Departmento′ ,≡, 0.5〉, w,−〉
aj,1 : 〈〈Persono, Persono′ ,≡, 1.0〉, l, +〉
aj,2 : 〈〈DepartmentHeado, Departmento′ ,�, 0.8〉, l, +〉
After generating their arguments, the matchers exchange their arguments with eachother. The matcher i sends to j its arguments ai,1 and ai,2, and vice-versa. i has apreference ordering w i l, while j has l j w. Having the complete set of arguments,the matchers generate their preferred extensions pi and pj . For both pi and pj , thearguments ai,1, aj,1 and aj,2 are acceptable: ai,1 and aj,1 are not attacked, while aj,2
successfully attacks ai,2 because both arguments have opposite values of h but aj,2
has highest strength than ai,2. So, the set of globally acceptable correspondences, A,contains both (1) and (2). It is the alignment associated with the objective consolidation.
Definition 11 (Alignment associated with an extension). Given an extension S in aSV AF , the alignment associated with this extensions is: A(S) = {c;∃〈c, v,+〉 ∈ S}.
However, this set is not consistent. Due to the fact that DepartmentHead is sub-sumed by Person in o, and Person and Department are disjoint concepts in o′, A isinconsistent as soon as there exists one Department.
4 Consistency-driven Argumentation
Resolving the inconsistency problem in alignment agreement has two possible alterna-tives: (a) express the inconsistency within the argumentation framework, as in [1, 4]; or
42
(b) deal alternatively with the logical and argumentative parts of the problem. Integrat-ing the logic within the argumentation framework seems a more elegant solution and itcan be achieved straightforwardly when correspondences are arguments and incompat-ible correspondences can mutually attack each others. However, this works only whentwo correspondences are incompatible. When the set of incompatible correspondencesis larger, the encoding is not so straightforward and may lead to the generation of anexponential amount of argument and attacks.
For that purpose, we define the consistency associated with an extension.
Definition 12 (Consistency). An extension S is said consistent iff its associated align-ment A(S) is consistent.
There are different ways to account for consistency in SVAF. The first one retainsonly consistent preferred extensions. However, the set of preferred consistent exten-sions may be empty. A fallback would be to consider maximal preferred consistentsub-extensions.
Definition 13 (Maximal preferred consistent sub-extensions). A consistent exten-sion S is a maximal preferred consistent sub-extension iff there exists a preferred ex-tension S′ such that S ⊆ S′ and ∀S′′; S ⊂ S′′ ⊆ S′, S′′ is not consistent.
There may be several such sub-extensions. Another approach, considered here, is towork on consolidations, i.e., the set of objective or subjective arguments.
Definition 14 (Maximal consistent sub-consolidations). A consistent extension S isa maximal consistent sub-consolidation of an (objective or subjective) consolidation S′
iff S ⊆ S′ and ∀S′′;S ⊂ S′′ ⊆ S′, S′′ is not consistent.
We propose a consistency-driven protocol that computes the maximal consistentobjective sub-consolidations. The algorithm removes the correspondences that intro-duce inconsistencies into the resulting alignment, for maintaining the coherence withinthe argumentation system. First, as in Section 3.1, the matchers compute their preferredextension from which the objective consolidation, O, is obtained. Based on O, the max-imal consistent sub-consolidations is then determined. It can be generalised to considersubjective consolidation or each preferred extension separately. If the objective con-solidation is consistent, then the algorithm returns it. If not, the maximal consistentsub-consolidation S is computed.
For computing S we have used the algorithm proposed by [14] which identifies theminimal sets of incoherent correspondences and removes them from the original align-ment. The algorithm is based on the theory of diagnosis, where a diagnosis is formedby the correspondences with lowest confidence degrees that introduce incoherence inthe alignment. It partially exploits incomplete reasoning techniques to increase runtimeperformance, preserving the completeness and optimality of the solution.
43
5 Experiments
5.1 Dataset, matchers and argumentation frameworks
The proposed approach is evaluated on a group of alignments from the conference trackof the OAEI1 2009 campaign. The data set consists of 15 ontologies in the domain ofconference organisation. They have been developed within the OntoFarm project2. Weuse the subset of these test cases where a reference alignment is available (21 align-ments, which corresponds to the alignment between 7 ontologies)3. We focus on equiv-alence correspondences, which are taken into account in the reference alignment, andfilter out subsumption correspondences.
We have chosen the alignments generated by the four best matchers that have par-ticipated in the 2009 OAEI conference track [7]: AMaker, Aflood, AMext and Asmov.
Each matcher has a SVAF and a private preference order, which is based onthe f–measure ordering for all matchers – AMaker (0.57), Aflood (0.52), AMext(0.51) and Asmov (0.47). The highest preferred value of each matcher is the valuethat it associates to its arguments. For instance, AMaker has as preference ordering:vamaker amaker vaflood amaker vamext amaker vasmov, while Asmov has theordering: vasmov asmov vamaker asmov vaflood asmov vamext.
For negative arguments (h = −), we use two different strength values. First, weconsider that the strength can vary according to the matcher quality (conformance withthe reference alignment). We assume that this strength is inversely proportional to theprobability that a false positive correspondence is retrieved by the matcher. Such proba-bility can be measured by the fallout of the alignment A, given the reference alignmentR. Then, we define str for the matcher m:
fallout(A,R) =| A \ R |
| A |, strm = 1 − fallout(Am, R)
Second, we use str=1.0, assuming that matchers strongly reject correspondencesthat they do not found (it could be the case when the information about the matcherquality is not available).
5.2 Results and discussion
We measure precision and recall of the maximal consistent sub-consolidation, S, withrespect to the reference alignments. First, we present the results from our approachand next we compare them with the results from each matcher. Figure 2 presents theresults from the objective consolidations, O, and from the maximal consistent sub-consolidation, S, for SVAFs with str = 1 and fallout-based str.
For SVAF with str = 1, argumentation (O) is sufficiently selective for generatingconsistent objective consolidations. We obtain high precision but low recall. This be-haviour is due to several reasons. First, we are using objective consolidations and only
1 Ontology Alignment Evaluation Initiative: http://oaei.ontologymatching.org/2 http://nb.vse.cz/˜svatek/ontofarm.html3 As in [7], the ontology Iasted is filtered out of our experiments because it causes reasoning
problems when combined with other ontologies. Thus, we have 15 test cases.
44
arguments present in every preferred extension are considered (what leads to an increasein precision). Correspondences being accepted by all matchers have high probability tobe consistent. Second, we use str = 1 for negative arguments (h = 1) and thus a truepositive (correct) correspondence with strength lower than 1.0 is successfully attackedby a false negative correspondence with strength 1.0 (what decreases the recall).
Using fallout-based str (Figure 2), we have an opposite behaviour. Argumentationis not able to filter out all inconsistent correspondences. We have low precision and highrecall. This occurs because negative arguments are not strong enough for successfullyattacking all positive arguments (including the incorrect ones). As a result, many cor-respondences are selected, what increases the probability for selecting inconsistent cor-respondences. When applying consistency checking, S, in average, precision slightlyincreases, while recall decreases. This effect is due the way the algorithm for removingthe inconsistencies works. An incorrect (but consistent) correspondence might causethe removal of all conflicting correspondences with lower confidence, and thus somecorrect correspondences are filtered out.
Fig. 2. SVAF with str = 1 and fallout-based str: objective consolidation (O) – intersection ofobjectively acceptable arguments for all audiences, without consistency-checking – and maximalconsistent sub-consolidation (S) – consistent subset of objectively acceptable arguments; andindividual results for each matcher.
Second, we compare the results from O and S with the results from each matcher.Figure 2 shows the matcher results with and without consistency checking. In the ma-jority of the test cases, the precision increases when filtering out the inconsistent corre-spondences, while recall decreases (in the case of Aflood, for some tests, the precisiondecreases while Amaker maintains its recall). As stated before, this is due to the fact thatsome correspondences are incorrect with respect to the reference alignment but consis-tent, as well as some correct correspondences are not included in the consistent setbecause together with some incorrect (but consistent) correspondences, they introduceinconsistencies into the set. Asmov is the only system able to check the consistency in
45
its alignments. In terms of f–measure, apart Asmov, consistency checking improves theresults from Amaker and Amext.
Comparing the results from SVAFs with the results from each matcher, for str=1(Figure 2), argumentation outperforms all matchers in terms of precision, but recall isbelow all matchers. For fallout-based str, we find an opposite behaviour. All matchersoutperform argumentation in terms of precision, but recall is better with argumentation.Looking for argumentation and consistency checking together, although consistencychecking slightly improves the precision, both precision and recall are below everymatcher. Consistency or argumentation improves results, while contrary to the intuition,we do not observe that the combination of both of these provide more improvements.
Following our (partial) experiments, we can observe that the behaviour of argu-mentation highly depends on the strength of the arguments. Argumentation is moreor less selective when using strong or weak strengths for negative arguments, respec-tively. Thus, an important issue in the argumentation model is related with the choiceof strengths of negative arguments.
Using logical consistency checking alone has positive effects in terms of f–measurefor the majority of matchers. On the other hand, combining argumentation and consis-tency checking slightly improves the precision, when argumentation is not sufficientlyselective for generating consistent alignments, but in terms of f–measure, this combina-tion has some negative effects. It is due particularly to the decrease in recall.
6 Related Work
Few ontology matching systems have been developed using semantic-based techniques.Examples of systems using some kind of logical verification are S-Match [9] and AS-MOV [11]. S-Match explores propositional satisfiability techniques (SAT) for generat-ing correspondences between graph-like structures. ASMOV semantically verifies thealignments for filtering inconsistencies. However, ASMOV lacks a well defined align-ment semantics and notions as correctness or completeness are thus not applicable [14].
In the field of alignment agreement based on argumentation, few approaches havebeen proposed. In [13], Bench-Capon’s model is used to deal with arguments that sup-port or oppose candidate correspondences between ontologies. Both Bench-Capon’sand SVAFs frameworks fail at rendering the fact that sources of correspondences oftenagree on their results, and that this agreement may be meaningful. [10] have adaptedthe SVAF in order to consider the level of consensus between the sources of the cor-respondences, by introducing the notions of support and voting into the definition ofsuccessful attacks. The work from [5] aims at identifying subparts of ontologies whichare sufficient for interpreting messages. This contributes to reduce the consumed time,at a minimal expense in accuracy.
In the field of alignment inconsistency, [15] and [12] considered correcting incon-sistent alignments. Revision is obtained exclusively by suppressing correspondencesfrom the alignment through minimising the impact of this suppression. In [15], the goalis to feed the consistent alignment back to a matcher so that it can find new correspon-dences. This process can be iterated until an eventual fix-point is reached. Similarly,[16] provides a revision operator by modifying one alignment between two ontologies
46
such that the result be consistent. Consistency and consequences are given by mergingboth ontologies and alignments within the same standard theory. Operators are providedbased on the notion of minimal conflict sets.
7 Concluding Remarks
We have defined consistency-driven argumentation for alignment agreement. This fillsa gap between argumentation-based matching and consistency-based alignment repairs.We have experimented our strategy on a set of alignments from expressive ontologies.The conclusion is that though theoretically grounded, the extra consistency step doesnot improve argumentation alone. At least in our experimental setting the argumentationprocess is incidentally able to provide near consistent extensions. We have analysed thefeatures of consistency checking and argumentation which cause this result.
Hence from these (partial) experiments we can conclude that applying inconsis-tency recovery and argumentation independently improves results, while using themtogether does not improve significantly the results. If this does not discard the validityof the approach, it reveals that it should not be applied without care, especially given itscomplexity.
Further study is required to know better in which context matching and argumenta-tion leads to inconsistency. One source of improvement would be to take into accountseveral such alignments between several ontologies (a network of ontologies). Indeed,these could raise inconsistency within networks of ontologies which would have to beconsidered as well.
Acknowledgements
We are grateful to Christian Meilicke for letting us use his consistent sub-alignmentsoftware.
References
1. L. Amgoud and P. Besnard. Bridging the gap between abstract argumentation systems andlogic. In Proceedings of the 3rd International Conference on Scalable Uncertainty Manage-ment, pages 12–27, Berlin, Heidelberg, 2009. Springer-Verlag.
2. L. Amgoud and C. Cayrol. On the acceptability of arguments in preference-based argumen-tation. In Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence,pages 1–7, San Francisco, California, 1998. Morgan Kaufmann.
3. T. Bench-Capon. Persuasion in practical argument using value-based argumentation frame-works. Journal of Logic and Computation, 13(3):429–448, 2003.
4. P. Besnard and A. Hunter. Argumentation based on classical logic. In I. Rahwan andG. Simari, editors, Argumentation in Artificial Intelligence, pages 133–152. Springer US,2009.
5. P. Doran, V. Tamma, I. Palmisano, and T. R. Payne. Efficient argumentation over ontol-ogy correspondences. In AAMAS ’09: Proceedings of The 8th International Conference onAutonomous Agents and Multiagent Systems, pages 1241–1242, Richland, SC, 2009. Inter-national Foundation for Autonomous Agents and Multiagent Systems.
47
6. P. Dung. On the acceptability of arguments and its fundamental role in nonmonotonic reason-ing, logic programming and n–person games. Artificial Intelligence, 77(2):321–357, 1995.
7. J. Euzenat, A. Ferrara, L. Hollink, A. Isaac, C. Joslyn, V. Malaise, C. Meilicke, A. Nikolov,J. Pane, M. Sabou, F. Scharffe, P. Shvaiko, V. Spiliopoulos, H. Stuckenschmidt, O. Svab-Zamazal, V. Svatek, C. T. dos Santos, G. A. Vouros, and S. Wang. Results of the ontologyalignment evaluation initiative 2009. In OM, volume 551 of CEUR Workshop Proceedings.CEUR-WS.org, 2009.
8. J. Euzenat and P. Shvaiko. Ontology matching. Springer, Heidelberg (DE), 2007.9. F. Giunchiglia, P. Shvaiko, and M. Yatskevich. S-Match: an algorithm and an implementation
of semantic matching. In Proc. 1st European Semantic Web Symposium (ESWS), volume3053 of Lecture notes in computer science, pages 61–75, Hersounisous (GR), 10-12 May2004.
10. A. Isaac, C. T. dos Santos, S. Wang, and P. Quaresma. Using quantitative aspects of align-ment generation for argumentation on mappings. In P. Shvaiko, J. Euzenat, F. Giunchiglia,and H. Stuckenschmidt, editors, OM, volume 431 of CEUR Workshop Proceedings. CEUR-WS.org, 2008.
11. Y. R. Jean-Mary, E. P. Shironoshita, and M. R. Kabuka. Ontology matching with semanticverification. Web Semantics: Science, Services and Agents on the World Wide Web, 7(3):235– 251, 2009. The Web of Data.
12. E. Jimenez Ruiz, B. Cuenca Grau, I. Horrocks, and R. Berlanga. Ontology integration usingmappings: Towards getting the right logical consequences. In Proc. 6th European semanticweb conference (ESWC), Heraklion (GR), volume 5367 of Lecture notes in computer science,pages 475–479, 2009.
13. L. Laera, I. Blacoe, V. Tamma, T. Payne, J. Euzenat, and T. Bench-Capon. Argumentationover ontology correspondences in mas. In AAMAS ’07: Proceedings of the 6th internationaljoint conference on Autonomous agents and multiagent systems, pages 1–8, New York, NY,USA, 2007. ACM.
14. C. Meilicke and H. Stuckenschmidt. An efficient method for computing alignment diagnoses.In RR ’09: Proceedings of the 3rd International Conference on Web Reasoning and RuleSystems, pages 182–196, Berlin, Heidelberg, 2009. Springer-Verlag.
15. C. Meilicke, H. Stuckenschmidt, and A. Tamilin. Reasoning support for mapping revision.Journal of logic and computation, 19(5):807–829, 2009.
16. G. Qi, Q. Ji, and P. Haase. A conflict-based operator for mapping revision. In Proc. 8thInternational Semantic Web Conference (ISWC), volume 5823 of Lecture Notes in ComputerScience, pages 521–536, 2009.
17. C. Trojahn, P. Quaresma, and R. Vieira. Conjunctive queries for ontology based agent com-munication in mas. In AAMAS ’08: Proceedings of the 7th international joint conference onAutonomous agents and multiagent systems, pages 829–836, Richland, SC, 2008. Interna-tional Foundation for Autonomous Agents and Multiagent Systems.
18. C. Trojahn, P. Quaresma, R. Vieira, and M. Moraes. A cooperative approach for compositeontology mapping. LNCS Journal on Data Semantic X (JoDS), 4900(0):237–263, 2008.
48
Alignment-Based Measure of the Distance between Potentially Common Parts of Lightweight Ontologies
Ammar Mechouche, Nathalie Abadie and Sébastien Mustière,
Institut Géographique National, Laboratoire Cogit, 73 Avenue de Paris, 94160 Saint-Mandé, France
{Ammar.Mechouche, Nathalie-f.Abadie, Sebastien.Mustiere}@ign.fr
Abstract. We propose in this paper a method for measuring the distance between ontologies susceptible to describe a common domain and for assessing the feasibility of their integration. This method is in two steps: the first step determines the potentially common parts of two ontologies, based on a prior alignment carried out between them. The second step computes the distance between these parts with regards to both their levels of detail and their structures, by exploiting the mappings contained in the alignment and adapting the Tree Edit Distance method. We limit our study here to lightweight ontologies, i.e., taxonomies represented in OWL1, the Ontology Web Language. This method was implemented and applied to real ontologies of the geographic domain. The results obtained so far seem significant.
Keywords: Distance between ontologies, Tree Edit Distance, Semantic Web.
1 Introduction
Measuring the distance between heterogeneous ontologies is useful for many applications: 1) retrieving ontologies on the web, e.g. finding an ontology to replace another [8], finding ontologies that can enrich other ones, finding people using same ontologies to create new collaborations, etc.; 2) ontology evolution, in order to know to what extent an ontology, especially its structure, has evolved; 3) ontology fusion and data integration, to know in advance if it may be possible to make joint studies on data described by heterogeneous ontologies. Ontology matching [7], which is addressed by many works and which consists in computing alignments between ontologies, i.e. determining correspondences between semantically related entities from heterogeneous ontologies, is a key idea enabling interoperability in the semantic web. It has brought solutions to some problems like finding ontologies for query translation. However, for the applications cited above we need to compute global similarity measures between ontologies. Indeed, for example, to enrich an ontology from another one, we need to know if they have close structures and if the second ontology is more detailed than the first one. An alignment between two ontologies does not allow knowing if these latter are complementary or not.
1 http://www.w3.org/TR/owl-features/
49
In the geographic domain, where data sources are annotated using heterogeneous ontologies [3][5][6][11], we mainly focus on assessing the similarity between ontologies based on three main criteria. The first similarity measure deals with the universes of discourse described by both ontologies: does the source ontology deal with the same domain as the target ontology or does it also provide additional knowledge about a related domain? For example, if we have a domain ontology describing landforms and vegetation, does the source ontology provide us, in addition to that, with knowledge about climate? The second similarity measure aims at
to assess whether they result from very different conceptualizations of the domain of interest or not: in other words, would it be difficult to communicate and exchange data with the community who produced this ontology? The third measure compares
ess precise than the target ontology. This aims at automatically determining whether available geodata sources have the appropriate thematic level of detail or not for a specific task. If we are looking for data describing buildings, we may need to make sure that they also explicitly describe more specific buildings such as cabins or huts.
We propose here a new method for computing the distance between potentially common parts of aligned lightweight [17] ontologies. Some works have been dedicated to evaluating the global similarity between ontologies [2], [9], [14-16]. However, their similarity values are difficult to interpret, since they do not measure the difference with regards to particular criteria such as the structure or the level of detail. Our method, however, uses a pre-computed alignment between two ontologies and provides the user with measures with regards to both their structures and levels of detail. This allows to more efficiently assessing differences between ontologies. Moreover, it computes distance between potentially common parts of ontologies, since two ontologies may be similar on one common thematic but quite different on another one. In fact, when source ontologies have common parts with the target ontology, their taxonomic structures must also be compared in order to evaluate to what extent they result from similar conceptualization of the domain of interest.
they are more or less precise than the target ontology. The remainder of this paper is structured as follows: section 2 presents our method
for measuring the distance between sub-parts of ontologies. Section 3 presents the results obtained with our method, and section 4 gives some perspectives to our work.
2 Proposed Method for Assessing Differences between Ontologies
Our method, first, uses a simple decision tree and the results of an alignment carried
of each ontology, which encompass a large number of mapped concepts. Once the important concepts are determined, the sub-parts of ontologies whose they are roots are compared. The comparison of ontologies (or ontology parts) here consists in computing the distance with respect to their structures and their levels of detail. To do that, we propose a Tree Edit Distance based method to compute the
50
distance between the structures of the compared ontology parts, and we exploit the results of the alignment performed between them to compute their levels of detail.
2.1 Determining the Important Concepts of two Aligned Ontologies
Our goal here is to determine the potentially common parts of two ontologies. To do that, we use existing tools to align the ontologies we want to compare and then we determine the parts of each ontology where the mappings are concentrated. Indeed, we believe that if two ontologies describe a common sub-domain then the concepts describing this sub-domain in both ontologies should generally be mapped together. We do not try here to improve the quality of the alignment, which is out of the scope of this paper. We suppose that the quality of the produced alignment is pretty good and the mappings it contains could be transformed into the following form:
ScoreDC OO ,,21
, which means that the concept C in the ontology O1 is mapped with the concept D in the ontology O2, and this mapping has a score confidence Score (between 0 and 1).
Fig. 1. The decision tree determining the important concepts of two aligned ontologies.
Before detailing our method, we consider that a concept is important if it encompasses a large number of mapped sub-concepts or if it encompasses a medium number of mapped sub-concepts but at the same time its depth in the ontology is small. In both cases, the important concepts define sub-parts of the ontology where the mappings are concentrated.
A first set of important concepts is determined thanks to the decision tree depicted in Fig. 1, which classifies each concept of the ontology in the class YES if it is an important concept or in the class NO if it is not . After that, additional rules allow to filter the first set of important concepts and to keep only the most significant ones.
The decision tree proposes three rules allowing determining important concepts: The first rule (leading to the YES in the continuous and bold rectangle) detects the
root of each ontology (the owl:Thing concept) as the important concept if the number of all mappings is high, since the root is the super class of any concept of the
51
ontology, and it encompasses all mapped concepts. On the examples shown on Fig. 2 the important concepts deduced by this rule are in continuous and bold circles.
The second rule (leading to the YES in the dashed rectangle) allows determining the important concepts which encompass a large number of mapped concepts. This rule does not allow to a concept C to be important if there are too few remaining concepts in the ontology that are not mapped and not sub-concepts of C. In this case it is better to consider the root of the ontology as the important concept. For example, on Fig. 2-(a1) the concept 3 could be an important concept if we suppose that all mapped concepts are its sub-concepts. However, since there remains only one concept (concept 2) in the ontology which is not its sub-concept it is better to consider only the root as an important concept in order to include this remaining concept. In Fig. 2(b1), however, there are two important concepts deduced by this rule (those with dashed lines). This is due to the fact that: 1) a large number of the mapped concepts are sub-classes of the concept 3 and consequently sub-classes of its parents ; 2) there remains a large number of concepts in the ontology which are not sub-concepts of 3 and which can constitute another part describing another thematic which is not common to the compared ontologies.
The third rule (leading to the YES in the shadow rectangle in the decision tree) allows determining the important concepts which encompass a medium number of mapped concepts and which are not deep in the ontology hierarchy. Indeed, in general, in an ontology, the distinction of the different described themes is made generally in its top level. In the example on Fig. 2(c1) the important concepts deduced by this rule are represented with shadow circles (2, 3 etc.). The important concepts are determined by this rule when the compared ontologies describe more than one common thematic.
Until now we determine a set of possible important concepts. For example, on Fig. 2(b1) and Fig. 2(c1) we have several important concepts which are deduced and we need to keep only the most significant among them, i.e. the more specific of them. To do this, we defined three additional rules allowing filtering these important concepts. They are tested in the following order, and only one of them is executed:
Filtering Rule 1: If we have only one important concept (the root), then keep it as
the important concept of the ontology. For example, on Fig. 2(a2) the concept 1 will be the important concept of the ontology; in other terms, the source ontology will be compared as a whole to the determined sub-parts of the target ontology.
Filtering Rule 2: If we have important concepts deduced by the second rule of the
decision tree (concepts in dashed circles), then we keep only the deepest among them as the important concept of the ontology. Indeed, in an ontology, the deeper the concepts are, the more they share common characteristics. In Fig. 2(b2), the important concept which will be kept is 3 since the Rule 1 will not be activated and the concept 3 is the deepest important concept deduced by the second rule of the decision tree. This way we obtain one important concept encompassing a large number of mapped concepts without including concepts describing other themes.
Filtering Rule 3: We keep all the deepest important concepts deduced by the third
rule of the decision tree (concepts in shadow circles). For example, in figure Fig.
52
2(c2), we keep only the important concepts 2 and 3 since they are the deepest ones. These concepts should be roots of two different themes described by the ontology.
Fig. 2. Possible cases for determining the important concepts.
2.2 Computing the Distance between the Ontology Parts
We consider here that two aligned ontologies O1 and O2 have the same structure if the subsumption relations (or is-a relations) between all mapped concepts are preserved in both ontologies. In other terms, the ontologies structures are the same if we do not consider the non mapped concepts. If there is one is-a relation between two mapped concepts C1 and C2 in O1, and this relation does not exists in O2 between the corresponding concept of C1 and the corresponding concept of C2, then we consider that the structures of O1 and O2 are different.
On Fig. 3(1) we have two ontologies with the same structure, but with different levels of detail. In this case, it is interesting to indicate to the user that these ontologies have similar structures, but that one is more detailed than the other one. This allows the user to decide, for example, to enrich the first ontology from the second one or to know if their fusion would be costly or not. In figure Fig. 3(2), however, both ontologies offer the same vocabulary, but their structures are different. Typically, in this case it would be costly to fusion or to combine these ontologies.
53
Fig. 3. Difference of levels of detail and structures between ontologies.
2.2.1 Comparing Structures
To compute the distance between the structures of two ontologies, we propose an adaptation of the Tree Edit Distance method [1], which is usually used to estimate the minimum effort which is necessary to transform an ordered tree into another one. We note that an ordered tree is a tree where the children of every node are ordered. The Tree Edit Distance method returns the minimum cost in terms of the number of operations (node insertion, node deletion and node renaming) which are necessary to transform one ordered tree into another one. Let us consider the example on Fig. 3(2). In order to transform tree1 into tree2 we need at least five operations (Fig. 4). So, transforming tree1 into tree2 costs 5.
In order to give a sense to the cost returned by the Tree Edit Distance method, we need to normalize it. To do this, we use the normalization formula proposed in [4] (formula (2)), where NC is the normalized cost considered as the distance between the two trees, C is the value returned by the Tree Edit Distance method, and |tree 1| and |tree 2| are the respective sizes (number of nodes) of the ordered trees tree1 and tree2.
21 treetreeCNC (2)
Fig. 4. Example of minimum operation number required to transform a tree into another one.
54
In the example on Fig. 4 the distance between tree 1 and tree 2 will be 5/(7+7) = 0.36, which means that similarity between tree 1 and tree 2 is 64%. In fact, the smaller the distance is, the closer the structures of ontologies are.
The adaptation of the Tree Edit Distance method is done as follows:
1. Every non mapped concept of each ontology is deleted and its mapped direct sub-concepts become direct sub-concepts of its closest mapped parent. Indeed, we are only interested in the structures formed by mapped concepts in each ontology. In Fig. 5, the concepts 8 and 9 in the ontology on the left are not mapped with concepts from the ontology on the right, so they are deleted and the concept 3 which is mapped becomes a direct child of the concept 1. The same reasoning is applied to the concept g.
2. The next step consists in relabeling the concepts in the second ontology by the labels of their corresponding concepts in the first ontology. If a concept C from the second ontology is mapped with one concept D from the first ontology, then C is renamed to D. For example, on Fig. 5, the concept a is mapped with the concept 1, then a is renamed to 1. If a concept C of the second ontology is mapped with several concepts from the first ontology, then it takes the name of the corresponding concept having the highest score of similarity. On the Fig. 5, the concept h is mapped with the concepts 6 and 7 from the first ontology, however the score of the mapping with 7 is higher, then h will be renamed to 7. If it had been mapped with several concepts from the first ontology having the same score of similarity, then the label of the most general subsuming concept of them would have been used for renaming. If no most general subsuming concept exists, then we would have chosen randomly the label of one of them for renaming. On the Fig. 5, the concept f is mapped with the concepts 5 and 6 from the first ontology with the same score, then f has to be renamed to 5, since 5 is more general than 6. We note that each concept from the first ontology is used at most one time to rename a concept in the second ontology.
3. The last step consists in ordering the obtained trees to allow using existing algorithms and tools for computing the Tree Edit Distance (which is our objective here). On the Fig. 5 we obtain two ordered trees whose distance was computed in the previous example (distance = 0.36).
Fig. 5. Transforming two aligned ontologies into two ordered trees.
55
2.2.2 Computing the Levels of Detail of two Aligned Ontologies
Two lightweight ontologies having close structures may have different levels of detail. We define here the level of detail of a concept C as the number of its sub-classes. For example, on Fig. 3(1) the concept 2 is more detailed in the ontology on the right than in the ontology on the left, since it has two sub-concepts in the ontology on the right when it has no sub-concept in that on the left. Now, we need to have an overall indication allowing to know which of the two aligned ontologies is more detailed. We propose to do this by averaging, for each ontology, the values of its mapped concepts. The following example illustrates this more clearly. Illustrating example. Let us consider the example on Fig. 3(1). To compute the level of detail of each ontology we create two vectors V1and V2, where V1 contains the level of detail values of each mapped concept from O1, and V2 contains the level of detail values of the corresponding concept in O2 of each mapped concept from O1. Then, the average value of V1 gives the level of detail of O1, and the average value of V2 gives the level of detail of O2. Thus, we obtain the following vectors:
321
ConceptsMapped
66.00021
1OLD
V
66.20262
2OLD
V
As we can see it, O2 is more detailed than O1.
3 Experiments and Results
The proposed methods were implemented and tested on real ontologies. The decision tree, determining the important concepts, was implemented in Java and using the Protégé OWL API2. The ontology structures comparison was also implemented as a Java program which uses the Protégé OWL API and reuses a Java implementation of the Tree Edit Distance method for ordered trees, available on the Web3 and described in [1]. The goal here is to determine which parts of two aligned ontologies are complementary, i.e. which parts are related by a large number of mappings and which have close structures but different levels of detail. We chose to test our methods on five ontologies describing geographic domains or domains close to geography, because our expertise in this domain allows us to better analyze the obtained results. The used ontologies are the followings: Building and Places ontology4: developed in United Kingdom, its purpose is to
describe the building feature and place classes surveyed by Ordnance Survey. Transportation ontology5: this ontology describes transportation-related
information in the CIA World Fact Book6.
2 http://protege.stanford.edu/plugins/owl/api/ 3 web.science.mq.edu.au/~swan/howtos/treedistance/ 4 http://www.ordnancesurvey.co.uk/ontology/BuildingsAndPlaces/v1.1/BuildingsAndPlaces.owl 5 http://reliant.teknowledge.com/DAML/Transportation.owl
56
Earth Realm ontology7: )8.
Hydrology9: this ontology is developed by Ordnance Survey to describe in an unambiguous manner the inland hydrology feature classes.
IGN ontology [12]: it is a bilingual ontology (French / English) which describes the topographic entities present in the geographic databases of the French Mapping Agency (IGN).
These ontologies are first pairwise aligned using the method proposed in the TaxoMap tool [10]. In order to determine the important concepts we defined the semantics of classification criteria used in the decision tree. So, we consider that the rate (percentage) of aligned concepts is high when it is superior to 80%. It is medium when it is comprised between 30% and 80%, and it is low when it is inferior to 30%. A concept of an ontology O is considered as deep if its depth in O is higher than a half of the depth of O itself. We note that, actually, these values are fixed intuitively, however we are working on in order determine them empirically. The obtained results with these values are summarized on Fig. 6.
Fig. 6. Important concepts obtained on real geographic ontologies.
We observe on Fig. 6 that most of the used ontologies describe the topography; many important concepts are in relation to the topography, which is a good result since most of the used ontologies really describe the topographic objects. Also, if we look at the size of each determined partition (whose root is an important concept) and the number of its mappings, we deduce that the important concepts detection is pretty precise. For example, the number of mappings between the Buildings and Places ontology and the IGN ontology equals 288. From one side, there is only one important concept deduced for each ontology (respectively Place and Artificial Topographic Feature). From the other side, the number of mappings included in the part whose
6 http://www.daml.org/ontologies/409 7 http://sweet.jpl.nasa.gov/1.1/earthrealm.owl 8 http://sweet.jpl.nasa.gov/guide.doc 9 http://www.ordnancesurvey.co.uk/ontology/Hydrology/v2.0/Hydrology.owl
57
Place is the root represents 89% of the total number of existing mappings, and the part whose Artificial Topographic Feature is the root contains all the mappings. So, the mappings are concentrated in the parts determined by our method.
The next step is the comparison phase. In order to obtain significant results we consider only mappings with a score higher than 0.90. We first computed the distance between the ontology parts structures. The obtained results are shown on Fig. 8.
Fig. 8. Results of the comparison of the ontology parts structures. The vertical axis indicates the computed Tree Edit Distance measures, and the horizontal axis indicates the compared ontology parts. Indeed, the numbers 1...9 refer to pairs of ontology parts that are compared. See Fig. 9.
Fig. 8 shows that there are some ontology parts that have more similar structures
than other ones. For example, the structure of the part of the Buildings and Places ontology whose the root is the important concept Topographic Object is very close to the structure of the part of the Hydrology ontology that has Topographic Object as a root. This is due to the fact that both ontologies are produced by the same institution, so with same conceptualization. The structure of the part of the Buildings and Places ontology whose the root is the important concept Place is, however, different from the structure of the part of the IGN ontology that has Artificial Topographic Feature as a root. In fact, in the metadata associated with Building and Places ontology, it is said
The rationale behind the Buildings and Places module is to provide a minimal set of definitions to maximise the abiliuty to reuse. As a result it contains a shallow hierarchy and minimal property restrictions his explains the difference in structures between its parts and parts of the IGN ontology which is very structured.
Finally, we compared the levels of detail (LD) of our ontologies using the method presented above. The results obtained are shown on Fig. 9.
Fig. 9. Result of the comparison of the ontology parts levels of detail (LD).
The results on the Fig. 9 are significant. For example, the Hydrology ontology is more detailed than the IGN ontology regarding the hydrographic features, and this can
58
be explained as follows: from one side the metadata associated with the Hydrology ontology say that the scope of this ontology includes permanent topographic features involved in the containment and transport of surface inland water of a size of 1 meter or greater including tidal water within rivers. From the other side, we know from the IGN databases specifications that the IGN databases, from which is built the IGN ontology, include only water surfaces larger than 7.5 meter. Another notable difference of levels of detail is between the Topographic Object part of the Hydrology ontology and the Topographic Object part of the Buildings and Places ontology. This result combined with the previous one telling us that the structures of these parts are very similar shows that these ontology parts are complementary and may help the user or a program to decide for example to enrich (for example thanks to an importation operation) the Topographic Object part of the Hydrology ontology from the Topographic Object of the Buildings and Places ontology.
4 Conclusion and Perspectives
Means for evaluating distance between ontologies seems to us important for decision making systems in the context of data integration, ontology fusion, ontology evolution and ontology retrieval on the web.
We have presented in this paper a new method for measuring the distance between lightweight ontologies. Our method differs from existing methods in several ways: 1) it exploits alignments between ontologies rather than assuming that ontologies share exactly the same vocabulary; 2) it does not compare whole ontologies but only the potentially common parts of them determined by our decision tree, in order to more efficiently assess the differences between ontologies; 3) finally, our method provides indications about the level of detail of each ontology and computes a distance between the ontologies structures by adapting the Tree Edit Distance method, which was not used in the past in this context to the best of our knowledge. The proposed method is implemented and tested on several real geographic ontologies, and the results obtained so far seem significant.
In the future, we plan to improve our method with respect to several aspects: 1) the ontology parts are determined with our decision tree which consists in detecting ontology parts where the mappings are concentrated, it would be interesting to compare and combine our decision tree with existing methods for ontology partitioning [13], in order to obtain better partitions. Moreover, our decision tree may be learnt using existing algorithms like the ID310 algorithm, particularly in order to automatically determine the semantics associated with each classification criteria; 2) our comparison method is actually restricted to lightweight ontologies, we plan to extend it to heavyweight ontologies in order include in our comparison procedures more complex constraints and relations between concepts modeled in domain ontologies; 3) another perspective of this work is to compare our method to similar ones [18] and to study the influence of different matching techniques on our distance measure; 4) finally, we plan to integrate to our method other information to better
10 http://en.wikipedia.org/wiki/ID3_algorithm
59
understand differences between ontologies, like metadata associated with ontologies, ontology utilization purposes, etc.
Acknowledgement
This work is partly founded by the French Research Agency through the GeOnto project ANR-O7-MDCO- , comparison and use of
.
References
1. Zhang K., Shasha D.: Fast algorithms for the unit cost editing distance between trees. Journal of Algorithms, 11(4), pp. 581-621, (1990)
2. Maedche A., Staab S.: Measuring Similarity between Ontologies. EKAW, 251-63, (2002) 3. Rodriguez M.A, Egenhofer M.J.: Determining Semantic Similarity Among Entity Classes
from Different Ontologies. IEEE TKDE, 15(2), p. 442-56. (2003) 4. Sager T., Bernstein A., Pinzger M., Kiefer C.: Detecting similar Java classes using tree
algorithms. International Workshop on Mining Software Repositories, 65-71, (2006) 5. Paul, M., Ghosh S.K.: An Approach for Service Oriented Discovery and Retrieval of
Spatial Data, Procedings International Workshop on Service Oriented Software Engineering (IW- 94, (2006)
6. Nambiar U., Ludäscher B., Lin K. and Baru, C.: The GEON portal: accelerating knowledge discovery in the geosciences, 8th ACM International Workshop on Web Information and Data Management, Aarlington, pp 83 90, (2006)
7. Euzenat J. and Shvaiko P.: Ontology Matching. Springer Verlag, p. 333, (2007) 8. Jérôme David, Jérôme Euzenat: Comparison between Ontology Distances (Preliminary
Results). International Semantic Web Conference, 245-260,(2008) 9. Wang J.Z., Ali F. and Srimani P.K: An Efficient Method to Measure the Semantic
Similarity of Ontologies. Advances in Grid and Pervasive Computing: p. 447-458, (2008) 10. Hamdi F., Zargavouna H., Safar, B. and Reynaud C.: TaxoMap in the OAEI 2008
alignment contest. Ontology Matching, p. 206-213, (2008) 11. Klien E.M.: Semantic Annotation of Geographic Information. Phd thesis, Institute for
Geoinformatics, University of Muenster. Muenster, Germany, (2008) 12. Abadie N. and Mustière S. s
spécifications de bases de données. Actes de SAGEO, Montpellier, (2008) 13. Stuckenschmidt H., Parent C. and Stefano Spaccapietra (Eds.): Modular Ontologies:
Concepts, Theories and Techniques for Knowledge Modularization. LNCS Vol. 5445 Springer, ISBN 978-3-642-01906-7, (2009)
14. Euzenat J., Allocca C., David J., d'Aquin M., Le Duc C. and Sváb-Zamazal, O.: Ontology distances for contextualisation. Report 3.3.4, IST NeOn IP, NeOn, (2009)
15. Ngan L., Soong A. and Hung L.: Comparing two ontologies. International Journal on Web Engineering and Technologies, 5(1): p. 48-68, (2009)
16. Mechouche A., Abadie N., Mustière S. : Mesure de la distance sémantique entre parties potentiellement communes à deux taxonomies, IC, France, (2010)
17. Óscar Corcho: Ontology based document annotation: trends and open research problems. IJMSO 1(1): 47-57 (2006)
18. P. Wang, B. Xu. Lily: Ontology Alignment Results for OAEI 2009. OM (2009)
60
Mapping the Central LOD Ontologies to PROTON Upper-Level Ontology
Mariana Damova, Atanas Kiryakov, Kiril Simov, Svetoslav Petrov
1 Ontotext AD, Tsarigradsko Chosse 135, Sofia 1784, Bulgaria { mariana.damova,naso,kivs,svetoslav.petrov}@ontotext.com
Abstract. Linking Open Data (LOD) facilitates the emergence of a web of linked data by publishing and interlinking open data on the web in RDF. One can explore linked data across servers by following the links in the graph. The LOD cloud has 203 datasets and more than 14 billion RDF triples (http://lod-cloud.net). This paper describes an approach to access these data by means of a single ontology, matched to the schemata describing several of the most common LOD datasets. They are presented in a reason-able view - FactForge (http://factforge.net) - the biggest and most heterogeneous body of factual knowledge on which inference is performed. Techniques of (a) making matching rules with “ontology expressions”, (b) adding new instances with inference rules, and (c) extending the upper level ontology with classes and properties are employed. They succeed to align ontologies designed according to different principles and displaying conceptual and structural mismatches.
Keywords: Linked Open Data, FactForge, PROTON, ontology matching, upper level ontology, semantic web, RDF, dataset, DBPedia, Freebase, Geonames.
1 Introduction
Linking Open Data (LOD) initiative [1] aims to facilitate the emergence of a web of linked data by means of publishing and interlinking open data on the web in RDF. One can explore linked data across servers by following the links in the graph in a manner similar to the way the HTML web is navigated. LOD cloud’s (figure 1) constantly increasing volume has a wealth of information which is of more than 14 billion RDF triples coming from a vast variety of data sources - 203 datasets. They are highly heterogeneous covering different subject domains with contribution from companies, government and public sector projects, as well as from individual Web enthusiasts. Accessing this wealth of data and making use of their full potential is still problematic. Linked data poses issues with respect to different dimensions: (a) open-world assumption of WWW data, combined with high complexity of reasoning even with OWL Lite, (b) some datasets are not suitable for reasoning, (c) publishing OWL datasets without accounting for its formal semantics. Linked data are generally unreliable as no consistency can be guaranteed. They are highly heterogeneous and hard to query. One way of accessing them is by using reason-able views [7] - an
61
approach for reasoning and management of linked data. A reason-able view (RAV) is an assembly of independent datasets, which can be used as a single body of knowledge with respect to reasoning and query evaluation. FactForge is such a reason-able view of the web of data.
Fig. 1. Linking Open Data cloud (LOD), [9].
It gathers 8 datasets from the LOD cloud - general knowledge (DBPedia, Freebase, UMBEL, CIA World Factbook, MusicBrainz), linguistic knowledge (Wordnet, Lingvoj), geographical knowledge (Geonames). FactForge is the biggest and most heterogeneous body of factual knowledge on which inference has been performed. It comprises an overall of 1.4 billion loaded statements, 2.2 billion stored statements and 10 billion retrievable statements. FactForge is developed as an evaluation case in the European research project LarKC [8] and is used as a testbed for different large scale reasoning experiments like WebPIE [11]. It is available as a free public service at http://factforge.com, offering the following access facilities: (a) incremental URI auto-suggest; (b) one-node-at-a-time exploration through Forest and tabulator linked data browsers; (c) RDF Search: retrieve ranked list of URIs by keywords; (d) SPARQL end-point. One can compose SPARQL queries with predicates from multiple datasets, as shown in figure 2.
Fig. 2. SPARQL query construction for FactForge datasets.
For example, the query
SELECT * WHERE { ?Person dbp-ont:birthPlace ?BirthPlace ; rdf:type opencyc:Entertainer ; ?BirthPlace geo-ont:parentFeature dbpedia:Germany . }
connects 4 datasets – DBPedia, OpenCyc, Geonames, and RDF. This powerful method to access the data from the LOD cloud has the drawback that one has to be familiar with all schemata and predicates of all datasets in FactForge in order to formulate the queries. It is even more difficult to automate the access to FactForge data and use the SPARQL end point in algorithms because of its heterogeneity. That
62
is why we envisaged a simplified way to access the data by providing an intermediary layer - a single ontology, as shown in figure 3. To do this, we chose to align the separate schemata of FactForge with the upper-level ontology – PROTON (the Base upper-level ontology (BULO)) [14].
Fig. 3. SPARQL query construction for FactForge datasets in the proposed approach.
The unified access point to FactForge using a single ontology as an interface to connect to all datasets in FactForge is designed to provide an easier and simpler access to the wealth of data, higher degree of interoperability and better integration of the datasets in FactForge. It allows obtaining information from many datasets via one single ontology schema. This unified access point has important applications such as semantic search and annotation using the entities from FactForge, semantic browse and navigation, querying FactForge in natural language, and many others. It should be clear however that the upper-level ontology does not cover the full diversity of the data in the datasets. Still, for specific fine-grained queries the original data schemata and ontologies should be used.
Thus, the main objective of our project was to build a foundational ontology to explore FactForge with a balanced class hierarchy and consistent three to four levels of depth. This implied extending PROTON to obtain optimal coverage of the rich data in FactForge. In addition, the structural and conceptual differences between PROTON and the schemata organizing the datasets of FactForge like DBPedia inspired the introduction of a method for extending FactForge datasets with new instances. So, the matching model of PROTON with FactForge schemata consists in a series of iterations of enrichments at conceptual and at data levels.
2 Approaches to Matching Ontologies
Ontology matching is a key interoperability enabler for the Semantic Web, as well as a useful tactic in some classical data integration tasks. It refers to the activity of finding or discovering relationships or correspondences between entities of different ontologies or ontology modules. Matching ontologies enables the knowledge and data expressed in the matched ontologies to interoperate. Distinct methods are employed to perform ontology matching. There are syntactic and semantic matching systems [3]. In the syntactic matching the relations are computed between labels at nodes, and they are evaluated as [0, 1]. In the semantic matching the relations are computed between concepts at nodes, and they are evaluated as set theoretic relations. The semantic matching discovers semantic relationships across distinct and autonomous generic
63
structures and recognizes relationships between matched entities, such as equivalence, subsumption, disjointness and intersection. When integrating two models, substantial difficulties may arise in transforming information from one model to the other in a heterogeneous context. Harmonising semantics is one approach for model integration by formal mapping between two domains. In this approach reference ontology is built to provide the link between the two models [3]. Except for the types of relationships that are matched between the ontologies, distinctions are made in the way the two initial ontologies are accessed. Thus, there are bidirectional and unidirectional matching methods. The bidirectional method ensures access to the two ontologies from the two ontologies, whereas the unidirectional method ensures access from one to the other ontology only [3]. Another difference in the matching methods is in the way the matching is done. There is manual and automated matching. Automated mapping is suitable for simple ontologies and simple matching tasks, where the exact accuracy of the matching is not of highest importance. In automatic matching structures that are being matched are labeled with natural language typically using WordNet. This is the vocabulary mapping. It consists in comparing Classes, Properties and Instances of two ontologies in a relation one to one. Automated matching competitions are carried out for several years now with tracks on different evaluation parameters [2], [4]. The benchmark track is run on one particular ontology dedicated to the very narrow domain of bibliography and a number of alternative ontologies of the same domain for which alignments are provided. The best result on this track of the 2009 matching competition is F-measure of 80% [4]. Extensive surveys of automated ontology matching methods can be found in [12], [13]. The main drawback of automated ontology matching systems is that they cannot cope with ontological heterogeneity. The fact is ignored that the classes and the properties may be described in different unrelated ontologies, thus the algorithms cannot discover hidden relationships that hold between unrelated entities. Mapping by hand is considered difficult, time consuming and too long, but it derives the most accurate results. Manual mapping is suitable when maximum quality of mapping is seeked for a small quantity of concepts.
Our adopted approach is unidirectional semantic manual alignment of PROTON and the ontologies of the selected datasets of FactForge.
3 The Data
This section describes the data on which the matching in our approach is being performed, e.g. PROTON (the Base upper-level ontology (BULO) [14]) and DBPedia, Freebase and Geonames of FactForge. They are ontologies built according to different design principles. PROTON is built according to the OntoClean method [5], [6] where, for example, type and role are distinguished. It consists in evaluating the ontology concepts according to Meta properties and checking them according to predefined constraints helping to discover taxonomic errors. Using the OntoClean methodology one can discover confusions between concepts and individuals,
64
confusions in levels of abstraction, e.g. object-level and meta-level, constraints violations, different degrees of generality.
The ontologies of FactForge datasets are made according to different methodologies. The ontologies of DBPedia and Geonames are data-driven. They provide structure and semantics to a large amount of entities in a shallow structure, but are however very different: DBPedia ontology includes many ad hoc predicates which appear in only one or several statements reflecting the variety of knowledge included in it. Geonames ontology has a concise conceptualization organized in very few well structured concepts and instances.
The upper level ontology – PROTON – is one side of the alignment process. An upper ontology is a model of the common objects that are applicable across a wide range of domains. It contains generic concepts that can serve as a domain independent foundation of other more specific ontologies. PROTON is built with a basic subsumption hierarchy comprising about 250 classes and 100 properties which provide coverage of most of the upper-level concepts necessary for semantic annotation, indexing, and retrieval.
DBPedia (http://dbpedia.org) is an RDFized version of Wikipedia. It is a collection of the structured information of Wikipedia, contained in its Infoboxes, represented in RDF and published on the Web. DBPedia ontology counts 24 first level concepts of very different degree of generality ranging from the philosophical concept of “event” through “person” and “place” to very specific concepts like “beverage”, “drug”, “protein”. Not all of DBPedia is comprised in the existing ontology. Many of the properties from the infoboxes are described separately as stand alone properties which pertain to ontological dimensions, but are not modelled in the ontology. Nevertheless some of these concepts are used in our alignment.
Freebase (http://freebase.com) is a large collaborative knowledge base, an online collection of structured data harvested from many sources, including individual wiki contribution. Freebase contains data from Wikipedia, Chemoz, NNDB, MusicBrainz and individually contributed data from its users. It has 5 million topics and no defined ontology. The entities described in this knowledge base are in structured predicate names, which reflect a hidden class hierarchy. Freebase has an overall of 19632 predicates with a structure of the predicate name in which the left most word denotes the subject domain of the property; the middle word denotes a class which is the domain of the property denoted by the last right most word, e.g.
������������� ������ � ������������������� ��������� ��������������� ������������
Geonames (http://geonames.org) is a geographic database that covers 6 million of the most significant geographical features on Earth. It contains over 8 million geographical names and consists of 7 million unique features whereof 2.6 million populated places and 2.8 million alternate names. All features are categorized into one out of nine feature classes and further subcategorized into one out of 645 feature codes. Geonames is integrating geographical data such as names of places in various
65
languages, elevation, population and others from various sources. All lat/long coordinates are in WGS84 (World Geodetic System 1984).
4 The Methodology
The project of building an intermediary layer between the heterogeneous data of FactForge and the end user requires matching of ontologies built according to different methods, e.g. data-driven ontologies and an upper-level ontology. This implies a translation from the one method to the other method. Further, the heterogeneity of the data in FactForge prompts the building of a unidirectional matching scheme, e.g. making FactForge accessible through PROTON predicates and entities, but not vice versa - PROTON through FactForge predicates and entities. The alignment was performed manually as the most suitable approach to find the correspondences of the small amount of upper-level concepts.
Our approach summarizes a method of matching ontologies with different methodological background – data-driven ontologies and an upper level ontology. The upper level ontology (PROTON) was chosen to be the basis for the mapping decisions, e.g. the representations of the other ontologies were translated into its model by (a) making matching rules with “ontology expressions”, (b) adding new instances with inference rules, and (c) extending the upper level ontology with classes and properties.
Thus, the adopted matching method includes:
• mapping of the concepts from PROTON to the concepts described in the datasets of FactForge, more precisely DBPedia, Freebase, Geonames
• assigning subsumption relations between entities and properties from FactForge to PROTON
• extending PROTON with classes and properties to obtain mapping at a conceptual level with FactForge
• using OWL class and property construction capabilities to represent classes and properties from FactForge and map them to PROTON classes
• extending FactForge with instances to account for the conceptual representations of the matching
The matching of the concepts and properties between DBPedia and PROTON and between Geonames and PROTON took place based on comparing the definitions of the concepts and their use. Respecting the commitment for unidirectional matching we have designed the rules with subsumption relations from FactForge to PROTON, as shown in the example below:
66
(a) dbp:Place rdfs:subClassOf ptop:Location .
(b) geo-ont:parentFeature rdfs:subPropertyOf ptop:subRegionOf .
But first, the upper level ontology PROTON was extended with new classes and properties. This was done after analyzing the content of the available data in DBPedia and Geonames with a result - a list of classes and properties which are represented within the data, and analyzing the structure of the current version of PROTON with respect to the new classes and properties. We obtained a classification of the new classes and properties using inheritance from already existing classes to the new ones. We have also used properties assigned to the new classes in order to structure them in a better way. Thus, we built a new version of PROTON with more classes and properties. Adding a new class or a new property in PROTON followed specific. A new class was added when the instances in FactForge formed a distinguishable group for which there was no concept description in PROTON. For example, DBPedia has instances for Fictional Characters, like Harry Potter, which are classified as Persons, the class FictionalCharacter was introduced in PROTON as a subclass of Person. A generic criterion for adding a new class to PROTON is the compliance with the principle of completeness of the ontology. This happens when for a given concept there are subconcepts represented in the ontology, but siblings of these concepts are missing. For example, if car and bicycle are subclasses of vehicle, but motorcycle is not, then we add motorcycle into the ontology.
To match Freebase predicates to PROTON the class construction capabilities of OWL have been used, to bind Freebase properties into classes and then match them to PROTON concepts as shown in the example (c) below:
(c)pfb:Location rdf:type owl:Restriction ; owl:onProperty <http://rdf.freebase.com/ns/type.object.type> ; owl:hasValue <http://rdf.freebase.com/ns/location.location> ; rdfs:subClassOf ptop:Location .
Here a class pfb:Location is created which is restricted to a Freebase type Location.
Another aligning method used is expression mapping. It consists in construction of classes on the basis of one of the ontologies, and mapping them to classes, or expressions of the other ontology, satisfying a relation of type many to many. For example, PROTON has a class Person and a class Profession. The subclasses of Person are Man and Woman and the subclasses of Profession are different professions, e.g. Architect, Teacher, etc. In DBPedia, Person is represented with the profession he exercises. Architect is a subclass of the class Person. Here we see a structural and conceptual difference between the PROTON model and the DBPedia
67
model with this respect. To perform the alignment we have adapted the DBPedia model to PROTON’s model in the mapping rule, as shown in figure 3.
Fig. 3. Mapping of concepts in ontologies designed according to different principles (PROTON, DBPedia).
Technically, the mapping rule looks like this:
(d) dbp:Architect rdfs:subClassOf [ rdf:type owl:Restriction ; owl:onProperty pupp:hasProfession ; owl:hasValue p-ext:Architect ] .
The professions are modeled as instances of the class Profession in PROTON, and the single entity of DBPedia is matched to an expression in PROTON which restricts the property hasProfession to the value of the profession of interest.
The method of expression matching is not universally applicable as described above. In some cases the expressions require a reference to instances which are not included in the datasets of FactForge. This triggered the next adopted aligning method - extending the dataset of FactForge with the necessary instances, ensuring their availability to cover the entire model of the chosen basic ontology - PROTON.
FactForge is loaded into BigOWLIM, the most scalable OWL engine (http://www.ontotext.com/owlim/) supporting light�weight and high�performance reasoning with inference based on OWL Horst. BigOwlim allows the definition of custom semantics via special rules and axiomatic triples which are exploited in the process of full materialisation performed during loading. This last mechanism was used to extend FactForge with new instances by adding inference rules to the built-in ruleset. The inference rules provide the insights on what triples have to be added into the repository. They are resolved at the time of loading of the datasets into the semantic repository. For example, the inference rule (e) below:
(e) p <rdf:type> <dbp-ont:PrimeMinister> ---------------------------------------
68
p <ptop:hasPosition> j j <pupp:hasTitle> <p-ext:PrimeMinister>
translates the DBPedia representation of someone holding a position of a Prime minister into PROTON representation. In DBPedia this is done with a type relation, whereas in PROTON this is a complex relation between a person holding a position with the title of Prime minister.
The translation of a single type relation in DBPedia can require more complexe representations, such as the ones given in example (f). Here the Freebase predicate government.us_president is represented as a person who holds a position in the US with the title president.
(f) a <fb:type.object.type> <fb:government.us_president>
----------------------------------------------------- a <rdf:type> <ptop:Person> a <ptop:hasPosition> y y <ptop:withinOrganization> <dbpedia:United_States> y <pupp:hasTitle> <p-ext:President>
Except for making the process of querying heterogeneous datasets easier, using one upper level ontology as an entry point to such data has another advantage. It allows to obtain information from many datasets via one single query. For example, one PROTON predicate covers three data driven predicates, e.g. PROTON locatedIntakes Freebase time.event.locations, and DBPedia place and location, as shown in the example (g) below.
(g) dbp:place rdfs:subPropertyOf ptop:locatedIn .
dbp-prop:location rdfs:subPropertyOf ptop:locatedIn .
<http://rdf.freebase.com/ns/time.event.locations> rdfs:subPropertyOf ptop:locatedIn .
This makes the exploration of FactForge richer and simpler, as a query with the single PROTON predicate will retrieve information with the three other predicates from the two different datasets. �
69
5 Results and Statistics
The outcomes of this work can be summarized as follows: (1) a new layer of unified semantic knowledge over FactForge was created by matching PROTON to FactForge schemata (2) we produced an original approach to providing similar layers to other datasets; (3) and developed a new version of PROTON ontology, which will be used in other projects. The extension of PROTON was governed by two main principles: (1) to provide coverage for the available data; and (2) to reflect the best approaches in the design of ontologies such as OntoClean methodology [5]. Table 1 shows statistics about the datasets of FactForge before and after the matching rules have been added to the semantic repository with full materialization performed. The alignment brought close to 800 million more statements and 50 million new entities available for exploration, while the matching rules cover 554 mapped classes and 103 mapped properties. The biggest number of mapped classes comes from the mapping of PROTON to Geonames’ feature codes (368). As far as PROTON enrichment is concerned, 166 new classes and 73 new properties have been introduced. They cover the classes which were identified during the analysis of the instance data in FactForge and their ontologies as described in section 4.
FactForge Initial State
FactForge with Alignments
Difference
Number of Statements 1,782,541,506 2,630,453,334 847,911,828 Number of ExplicitStatements 1,143,317,531 1,942,349,578 799,032,047 Number of Entities 354,635,159 404,798,593 50,163,434
Table 1. Statistics of FactForge
The adopted method was tested on 27 evaluation SPARQL queries selected to cover different domains, e.g. public administration, military conflicts, art and entertainment, business, medicine and to use multiple datasets from FactForge. Table 2 presents an example of an evaluation query. It is about cities around the world which have “Modigliani art works”. This query is considered the ultimate test for the Semantic Web [10]. To our knowledge FactForge is the only engine capable of passing this test. The right column of the table gives the query written with PROTON predicates only. It is simpler and more intuitive than the FactForge standard one as the mapping has put all FactForge location predicates into one PROTON predicate. The number of results returned with PROTON query and with FactForge standard query are the same, presented in a slightly different way. This proves the validity of the approach.
FactForge – Standard FactForge - PROTON
PREFIX fb: <http://rdf.freebase.com/ns/> PREFIX dbpedia: <http://dbpedia.org/resource/> PREFIX dbp-prop: <http://dbpedia.org/property/> PREFIX dbp-ont: <http://dbpedia.org/ontology/> PREFIX umbel-sc: <http://umbel.org/umbel/sc/> PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX ot: <http://www.ontotext.com/>
SELECT DISTINCT ?painting_l ?owner_l ?city_fb_con ?city_db_loc ?city_db_cit
WHERE { ?p fb:visual_art.artwork.artist dbpedia:Amedeo_Modigliani ; fb:visual_art.artwork.owners [ fb:visual_art.artwork_owner_relationship.owner ?ow ] ; ot:preferredLabel ?painting_l. ?ow ot:preferredLabel ?owner_l .
PREFIX dbpedia: <http://dbpedia.org/resource/> PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX ot: <http://www.ontotext.com/> PREFIX ptop: <http://proton.semanticweb.org/protont#> PREFIX ploc: <http://proton.semanticweb.org/protonl#> PREFIX p-ext: <http://proton.semanticweb.org/protonue#>
SELECT DISTINCT ?painting ?owner ?city
WHERE { ?p p-ext:author dbpedia:Amedeo_Modigliani ; p-ext:ownership [ ptop:isOwnedBy ?ow ] ;
ot:preferredLabel ?painting . ?ow ot:preferredLabel ?owner . ?ow ptop:locatedIn [ rdf:type ploc:City ; ot:preferredLabel ?city]. }
70
OPTIONAL { ?ow fb:location.location.containedby [ ot:preferredLabel ?city_fb_con ] }
OPTIONAL { ?ow dbp-prop:location ?loc. ?loc rdf:type umbel-sc:City ; ot:preferredLabel ?city_db_loc }
OPTIONAL { ?ow dbp-ont:city [ ot:preferredLabel ?city_db_cit ] } }
����������Table 2. Modigliani Test Query
In cases where several FactForge predicates are matched to a single PROTON predicate, like the location predicates mentioned earlier in the paper, the PROTON queries return more results than FactForge – Standard queries. Thus, the advantages of the approach to have a single access point to the Linked Open Data (LOD) cloud are twofold: they provide access by simpler queries and they provide leveraged query results.
5 Future work
We envision in the future building a two level intermediary layer to access FactForge and then LOD cloud mapping PROTON to UMBEL (http://www.umbel.org/documentation.html) – “a lightweight subject concept reference structure for the Web” with about 20 000 subject concepts based on OpenCyc (http://www.cyc.com/opencyc/). We intend to cover more datasets from the LOD cloud, and to experiment with the balance between the data from the LOD and FactForge datasets and the ontological schemata describing them. �
References
1. Christian Bizer, Tom Heath, Tim Berners-Lee. Linked Data – The Story so Far. In: Heath, T., Hepp, M. and Bizer, C. (eds.) Special Issue on Linked Data, International Journal on Semantic Web and Information Systems (IJSWIS), http://linkeddata.org/docs/ijswis-special-issue, (2009).
2. Caterina Caracciolo, Jérôme Euzenat, Laura Hollink, Ryutaro Ichise, Antoine Isaac, Véronique Malaisé, Christian Meilicke, Juan Pane, Pavel Shvaiko, Heiner Stuckenschmidt, Ondˇrej Šváb-Zamazal, and Vojtˇech Sváte, Results of the Ontology Alignment Evaluation Initiative 2008, (2008).
3. Mohamed El-Mekawy and Anders Östman. Semantic Mapping: an Ontology Engineering Method for Integrating Building Models in IFC and CITYGML. In Proceedings of the 3rd ISDE Digital Earth Summit 12-14 June, 2010, Nessebar, Bulgaria, (2010).
4. J. Euzenat, A. Ferrara, L. Hollink, A. Isaac, C. Joslyn, V. Malaisé, C. Meilicke, A. Nikolov, J. Pane, M. Sabou, F. Scharffe, P. Shvaiko, V. Spiliopoulos, H. Stuckenschmidt, O. Šváb-Zamazal, V. Svátek, C. Trojahn dos Santos, G. Vouros, S. Wang: Results of the Ontology Alignment Evaluation Initiative 2009. In Proceedings of the 4th Ontology Matching Workshop at ISWC, (2009).
5. Nicola Guarino and Christopher Welty. “Evaluating Ontological Decisions with OntoClean.” Communications of the ACM, 45(2): 61-65 (2002).
71
6. Michael Grüninger and Mark Fox. Methodology for the Design and Evaluation of Ontologies. Proceedings of IJCAI95's Workshop on Basic Ontological Issues in Knowledge Sharing (1995).
7. Atanas Kiryakov, Damyan Ognyanoff, Ruslan Velkov, Zdravko Tashev, Ivan Peikov. LDSR: Materialized Reason-able View to the Web of Linked Data. In: Proceedings of OWLED 2009. Chantilly, USA, 23-24 October 2009 (2009).
8. LarKC project. The Large Knowledge Collider - http://www.larkc.eu/, (2010). 9. Linking Open Cloud Diagramme, as of September 2010, http://lod-cloud.net/. 10. Richard MacManus. The Modigliani Test: The Semantic Web’s Tipping Point.
http://www.readwriteweb.com/archives/the_modigliani_test_semantic_web_tipping_point.php (2010).
11. Vassil Momchev, Matthias Assel, Alexey Cheptsov, Barry Bishop, Luka Bradesko, Christoph Fuchs, Georgina Gallizo, Spyros Kotoulas, Gaston Tagni D5.5.3 Report on platform validation and recommendation for next version, LarKC EU-IST-2008-215535, (2010).
12. P. Shvaiko, J. Euzenat: Ten Challenges for Ontology Matching. In Proceedings of ODBASE, (2008).
13. P. Shvaiko, J. Euzenat: A Survey of Schema-based Matching Approaches. Journal on Data Semantics, 2005.
14. Ivan Terziev, Atanas Kiryakov, Dimitar Manov. „D.1.8.1 Base upper-level ontology (BULO) Guidance” Deliverable of EU-IST Project IST – 2003 – 506826 SEKT (2005).
72
Ontology Alignment in the Cloud
Jurgen Bock, Alexander Lenk, and Carsten Danschel
FZI Research Center for Information Technology, Karlsruhe, Germany{bock,lenk,daenschel}@fzi.de
Abstract. The problem of ontology alignment is prominent for applica-tions that operate on integrated semantic data. With ontologies becom-ing numerous and increasingly large in size, scalability is an importantissue for alignment tools. This work introduces a novel approach forcomputing ontology alignments using cloud infrastructures. An align-ment algorithm based on particle swarm optimisation is deployed on acloud infrastructure, taking advantage of its ability to harness parallelcomputation resources. The deployment is done with a focus on parallelefficiency, taking into account both communication latency and compu-tational inhomogeneity among parallel execution units. Complementingprevious experiments showing the effectiveness of the alignment algo-rithm, this paper contributes an experiment executed “in the cloud”,which demonstrates the scalability of the approach by aligning two largeontologies from the biomedical domain.
1 Introduction
Ontology alignment is a problem prominent in semantic applications, the seman-tic web, and the linked data web. It is based on the observation that ontologies1
are heterogeneous models, often representing a similar or equal domain of in-terest, and hence have a certain overlap. Accordingly an ontology alignmentis defined as a set of correspondences between ontological entities, i.e. classes,properties, and individuals, of two ontologies.
Two examples of where ontology alignment plays an important role are thelinked data web [2] and the biomedical domain [21]. These examples also demon-strate the necessity for alignment tools to be able to deal with very large on-tologies, which currently constitutes a problem for most alignment algorithms.Furthermore, it can be observed that ontologies evolve gradually, i.e. changestypically occur by adding / removing / modifying single entities while the largestpart of the ontology remains unchanged. Thus, due to the typically slow changes,alignments do not need to be recomputed completely each time, but rather incre-mentally maintained and adjusted according to the evolving ontologies. Further-more, in typical information systems that utilise alignments, the refresh period
1 There is some disagreement among semantic web and linked data web communitiesabout whether data sources in the linked data web can be called ontologies. In thispaper the term ontology is used to cover all types of semantic data sources.
73
for alignments is not time-critical, i.e. alignments do not need to be recomputedad-hoc, but can be adjusted offline, e.g. in nightly alignment jobs.
This paper addresses the scalability problem of ontology alignment by util-ising cloud computing as a massively parallel computation infrastructure. Thealgorithm to be deployed in the cloud is an alignment algorithm based on particleswarm optimisation (PSO) [4]. A PSO-based approach to the ontology alignmentproblem has several characteristics that meet the aforementioned observations:
1. The algorithm provides a meta-heuristic, which is independent of the ob-jective function to be optimised. Hence it is straightforward to exchange oradapt the objective function according to the alignment scenario at hand.
2. The algorithm works incrementally, which has two beneficial effects: Firstly,the algorithm can be interrupted at any time providing the best alignmentthat has been discovered so far. Secondly, the algorithm can be provided withan initial (partial) alignment as a start configuration to be refined. This inparticular allows for alignment evolution of gradually changing ontologies.
3. The algorithm is inherently parallelisable allowing for efficient execution onparallel computing infrastructures.
The first two issues are inherent to the application of the PSO meta-heuristic.However, the usability of the approach depends on an efficient deployment onparallel computing infrastructures. Computation and adjustment of alignmentson an irregular basis usually does not justify the acquisition of large parallelhardware infrastructures.
Cloud computing infrastructures are scalable and can be used for data in-tensive parallelisable jobs, as it has successfully been shown e.g. in the field ofbio-informatics [20]. Recently, cloud computing has also been recognised as apromising technique to provide scalability for web data processing [18]. The useof cloud infrastructures provides the possibility to access a large number of com-puting resources in a convenient way without the need of operating one’s own,expensive data centre. Many offerings by cloud providers such as Amazon WebServicesTM (AWS)2 are based on a pay-per-use pricing model with an hourlyrate. This allows for flexible use of computation resources, since accessing largeamounts of computation power for a short period of time comes at the samecosts as using just a few resources for a long period of time.
The remainder of this paper is structured as follows. In Sect. 2 relevantbackground knowledge and related work is presented. Section 3 introduces theapproach of ontology alignment by PSO and demonstrates its parallel scalability.Justified by these insights, Sect. 4 discusses the design and implementation ofa cloud-based deployment of the formerly introduced approach. Experimentalresults are presented in Sect. 5 demonstrating both the effectiveness by referringto previous benchmarks, and the scalability by using a real-world biomedicalalignment scenario. Section 6 summarises the contributions and provides anoutlook on future work.
2 http://aws.amazon.com
74
2 Foundations
This section introduces ontology alignment, particle swarm optimisation (PSO),and cloud computing in more details.
2.1 Ontology Alignment
An ontology alignment is defined as a set of correspondences between ontologicalentities [10], i.e. classes, properties, and individuals, of two ontologies. In thisapproach, an entity can correspond to at most one entity of the other ontology,which denotes what Euzenat and Shvaiko call an ?:? alignment [10]. Ontologyalignment detection can be perceived as an optimisation problem [4], where thetask is to find the optimal alignment of two ontologies w.r.t. the criteria denotedin terms of an evaluation function. These criteria can be based on several basicmatching techniques [10, Chap. 4] or other measures, respecting domain specificand modelling language specific characteristics of the ontologies to be aligned.
A representative overview of the state-of-the-art in ontology alignment isgiven by Euzenat and Shvaiko [10, Chap. 6] and by the yearly Ontology Align-ment Evaluation Initiative (OAEI) [9]. Most systems focus on the improvementof alignment quality, whereas the scalability problem has attracted interest onlyrecently. As the OAEI participation shows, only few systems are capable of deal-ing with large ontologies [9, Sect. 10].
2.2 Particle Swarm Optimisation (PSO)
PSO is a biologically and socioculturally-inspired meta-heuristic [12, 13, 8], orig-inally proposed in 1995 by Kennedy and Eberhart [12]. It has become a majorresearch field since then.
PSO algorithms use a population of particles to find the optimal parameterconfiguration with respect to one or more objective functions. Each particlerepresents a candidate solution. A particle’s fitness is evaluated using objectivefunction(s). During initialisation, each particle in the swarm is assigned a randomposition in the parameter space. A PSO algorithm runs iteratively, where in eachiteration, each particle adjusts its position in the parameter space by adding avelocity vector to its current position. These particle updates can be performed inparallel. The velocity vector for a particle is determined by the particle’s previousvelocity (inertia), the best position that has been visited by any neighbour ofthe particle so far (social component), and the best position, the particle itselfhas visited so far (cognitive component). Regarding the social component, severalmodels of defining particle neighbourhoods have been analysed. Such models arecalled social network structures or topologies. A PSO algorithm is called gBest(global best) PSO, if the neighbourhood of a particle consists of the whole swarm.Thus the gBest PSO is a special case of the lBest (local best) PSO, where theneighbourhood of a particle comprises only a subset of the swarm [8, Chap. 12].
It is important to note, that the particle swarm search heuristic works withoutany knowledge about the objective function(s). Hence an objective function isreplaceable and adjustable to arbitrary optimisation problems.
75
2.3 Cloud Computing
Cloud computing is a new paradigm that has been evolving over the last fewyears. Most definitions of cloud computing have in common that cloud computingofferings can be categorised using the “Everything as a Service” (XaaS) model. Amore detailed view of this model is the “Cloud Computing Stack” [16]. Accordingto the XaaS model the three main service classes are “Software as a Service”(SaaS), “Platform as a Service” (PaaS), and “Infrastructure as a Service” (IaaS).While SaaS offerings usually provide an interface directly to the end user byproviding a Web service interface or a graphical user interface (GUI) the PaaSand IaaS offerings can be used by software architects to build new SaaS serviceson top of them. PaaS offerings usually provide a platform where the softwaredeveloper can deploy the new services [16].
The IaaS offerings at the “Basic Infrastructure Services” level, give the de-veloper full control over the servers that are running his software. At this levelthe user can deploy new machines using Web service technologies. This offersthe power and flexibility to work on a machine level without running one’s owndata center and so having convenient access to new resources. Offerings such asAmazon EC2 give users the opportunity to automatically deploy hundreds ofvirtual machines within minutes. Thus, it is possible to build highly scalable ap-plications that can scale up and down in short periods. One famous example ofa successful cloud offering is Animoto [1]. The Animoto application transformsmusic files and photos into small video slide shows. After offering their service tousers on a social network the demand of virtual machines went from about 40 to3500 [15]. This scalability was only possible by designing the Animoto softwareas a distributed algorithm deployed on Amazon EC2.
Another interesting feature of cloud offerings is the typical pay-as-you-gopricing model. This means that users only pay for the resources they are reallyusing. Having such a pricing model it makes no difference if one single serveris running for 10 hours or if 10 servers are running for just one hour. The NewYork Times used this pricing scheme when they built their “TimesMaschine”.By using Amazon EC2 they were able to convert their whole archive (4TB ofscanned TIFF files), spanning the years 1851-1922, to web documents within24 hours and total costs of US$ 890 [11].
In the context of semantic technologies and the semantic web, cloud comput-ing technologies have been successfully applied, mainly for RDF storage [22, 19,18], querying [18], and materialisation of RDF/OWL knowledge bases [24, 23].
3 Ontology Alignment by Particle Swarm Optimisation
Considering ontology alignment as an optimisation problem, a novel discretePSO algorithm (DPSO) has been developed to find an optimal alignment of twoontologies [4]. The algorithm has been implemented under the name MapPSO3
and is based on a DPSO algorithm by Correa et al. [6], which implements efficientattribute selection in data mining classification tasks.
3 http://mappso.sourceforge.net
76
3.1 Algorithm
In the case of ontology alignment, each particle represents a valid candidatealignment. With a swarm of particles being initialised randomly, in each itera-tion, each particle evaluates the alignment it represents via an objective functioncomprising various basic matching techniques [10, Chap. 4] as well as the size ofan alignment. Typically there is only a partial overlap of ontologies, so the align-ment algorithm strives for finding the largest alignment, i.e.maximum number ofcorrespondences, of high quality. Base matchers can evaluate a correspondenceaccording to different characteristics, such as lexical or linguistic similarity ofentity labels or comments. Moreover, correspondences do not necessarily haveto be evaluated in isolation, e.g. a correspondence of two classes can get a betterevaluation if a correspondence of their superclasses is also part of the alignmentrepresented by this particle. As discussed in Sect. 2.2 the objective function andthus the base matchers are replaceable and hence can be adapted to particularalignment scenarios. For instance in biomedical ontologies specific OBO4 [21]annotations can be exploited to evaluate the similarity between entities.
Similar to the previous work by Correa et al. [6] and the original binary PSOalgorithm by Kennedy and Eberhart [14], particle movements are determinedby proportional likelihoods, since the traditional notion of velocity vectors is notapplicable for binary search spaces, as it is the case for ontology alignment. Thelikelihood for a correspondence to be contained in a particle in the subsequentiteration is increased, if the correspondence exists in the global best (social com-ponent) or personal best (cognitive component) particle configuration.
This update mechanism allows for a guided convergence of the swarm to aglobal optimum, which represents the best alignment with respect to the basicmatching techniques specified as objective function. Moreover, the size of eachparticle varies during the execution of the algorithm influenced by the size ofthe global best alignment. This causes the algorithm to search for the optimumsize of the alignment, i.e. partial overlap, as well.
3.2 Parallel Efficiency
A detailed theoretical discussion of the correlation between population size andnumber of iterations with respect to the swarm convergence goes beyond thescope of this paper. Figure 1 illustrates an empirical analysis of the convergencetowards the optimum5 when aligning small ontologies of the OAEI benchmarkdataset. The figure shows clearly that a larger number of particles results in fasterconvergence by reducing the number of required iterations and thus reducingwall-clock runtime. Where it takes about 225 iterations for a swarm of 8 particlesto reach an alignment of high quality, the same result is achieved in only 90iterations by a swarm of 64 particles. Using only 2 particles does not reach anequally good result within 300 iterations.
4 Open Biomedical Ontologies5 Note that the optimum depends on the chosen base matchers and thus not necessarilyhas to be 0.
77
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 50 100 150 200 250 300
Qua
lity
of g
loba
l bes
t Alig
nmen
t
Iteration
2 particle8 particle
32 particle64 particle
Fig. 1. Convergence of the PSO algorithm for different population sizes using test case101 of the OAEI benchmarks. Alignment quality is represented as the fitness of theglobal best particle, where a smaller value represents better alignment quality.
4 Deployment in the Cloud
To obtain the required scalability, the PSO-based alignment algorithm has beenported to Amazon Web ServicesTM (AWS), the IaaS cloud service of Amazon R©.AWS is one of the biggest IaaS providers with a large and active community,and is representative for any other IaaS provider in the context of this work.
The deployment has been realised using a server-worker pattern. Severalworkers evaluate and update several local particles each, while they are managedby a central server. Each worker determines the local best alignment amongthe results of its particles and sends it to the server. The server determines theglobal best alignment, then broadcasts it and synchronises the workers. Exchangeof information between server and workers is realised by the Amazon SimpleQueue Service (SQS)6 or via TCP/IP. The exchange of concrete particle statesis realised by the Amazon Simple Storage Service (S3)7 for reasons of scalability,reliability, and parallel access.
4.1 Challenges
Deploying the algorithm on virtual machines connected by a network bears twomain challenges. Firstly, the communication latency is much higher when com-municating via network than via main memory. Hence finding and broadcastingthe global best particle leads to a higher communication overhead, which slowsdown the algorithm and creates unwarranted costs.
6 https://aws.amazon.com/sqs/7 https://s3.amazonaws.com/
78
Secondly, the computation times of workers in a particular iteration differ.This difference occurs mainly for two reasons: The unpredictable performanceof the virtual environment, and the varying particle sizes. The performance ofthe virtual machine depends on its mapping to real hardware. For example a“noisy” neighbour, i.e. another program sharing the same real hardware viaa different virtual machine, can slow down network communication. In turn, avirtual machine can utilise more computing power if there are no or only inactiveneighbours. This results in unbalanced performance of the virtual machines. Therandom initialisation of particles and their different sizes add to this discrepancyin computation time. A small particle needs less computation time than a bigparticle, therefore a worker with smaller particles will require less runtime periteration. The random initialisation of particles with different sizes is necessaryto search for the optimal alignment size and thus be able to identify partialoverlaps of ontologies. This computation time discrepancy causes fast workersto idle while waiting for the slower workers and thus decreases parallel efficiency.
4.2 Increasing Parallel Efficiency
The challenges of deploying the algorithm to a cloud-based infrastructre identi-fied have been addressed and solutions are proposed as follows.
Addressing Latency. Amazon SQS was used as a means for communicationbetween server and workers. Amazon advertises SQS as reliable, scalable, andsimple. However, it turned out that SQS has high latency. For reducing thenetwork latency, direct communication has been implemented using the TCP/IPprotocol. Apart from reduced latency, this also resulted in a reduction of theadditional communication overhead caused by multiple workers.
The latency is reduced further by only sending particle states when necessary,i.e. when a better particle state has been found. To achieve this a worker sendsthe fitness value of its local best particle to the server and writes the particlestate itself into the S3 database. The server then broadcasts the global bestfitness and the according database location to all worker instances.
The number of read operations is minimal, since in the PSO topology used,each particle must know about the global best. In case the global best does notchange, no worker has to read a particle state. The (unlikely) worst case in termsof communication latency happens if every worker finds a new best alignmentand thus writes its particle state before it has to read a new one found by anotherworker in the same iteration.
Addressing Runtime Discrepancy. The effect that workers require differentruntimes for particle fitness computation can be minimised by particle pooling,i.e. having each worker instance computing multiple particles. Using few particlesper worker results in high runtime variance between workers caused by differentparticle sizes and the resulting uneven workload. Using more particles per worker,i.e. the workload for each worker being the sum of workloads contributed by each
79
of its particles, averages the runtime because a rather uniform distribution ofsmall and big particles per worker is expected. Using a multi-particle approachalso increases parallel efficiency due to the increased overall runtime required toevaluate and update several particles. By increasing the runtime the proportionof time used for communication decreases and thus parallel efficiency is increased.
Using asynchronous particle updates is another way to compensate the run-time discrepancy. When using synchronous particle updates every worker has towait for the slowest worker and thus is wasting computation time. In the asyn-chronous communication mode workers keep on evaluating and updating theirparticles until they find a particle state that is better than the global best theyknow about. They send this particle state to the server and continue. The serverbroadcasts the new particle state if it is better than the global best known bythe server. Preventing workers to idle drastically increases parallel efficiency.
Introducing an asynchronous particle update strategy has an effect on theunderlying particle swarm meta-heuristic. Having not all particles exchanginginformation at the same time step has a similar effect than changing the socialnetwork structure of the PSO from a star topology to a cluster topology8 [8,Chap. 12], which results in fast communication between particles on the sameworker compared to the communication between workers themselves. A clustered(lBest) topology in general results in slower convergence but better robustnesscompared to the star (gBest) topology [8, Chap. 12].
Furthermore, a loss of information can be observed compared to the syn-chronous update mechanism. This is due to the fact, that particles compute it-erations with possibly outdated information about the global best, which mightbe available on a worker that is still computing another particle. This problemhas been analysed by Lewis et al. [17]. Their investigations revealed that theinformation loss can be compensated by the increased number of computationsthat are possible due to the reduced waiting time.
The beneficial effect on the runtime for an alignment is reflected in Fig. 2,showing runtime behaviour for two medium sized ontologies for both synchronousand asynchronous particle updates. As expected, using an asynchronous commu-nication mode does not have any effect if only a single worker is used. However,for 16 workers that need to communicate their local best results in each iteration,a clear runtime improvement of almost 50% can be observed.
One further effect resulting from the asynchronous particle updates is thatworkers hosting mainly small particles can complete an iteration more quicklythan those hosting mainly large particles. Therefore small particles tend to com-pute more iterations and thus influence the swarm stronger than it would be thecase in the synchronous mode. This is beneficial to the overall runtime of thealgorithm, because the average size of particles is smaller and therefore iterationsare faster.
8 While in a star topology, every particle shares information with every other particle,the cluster topology allows only groups of particles to communicate with each other,while the groups themselves exchange information only via dedicated particles, whichare part of two clusters.
80
Fig. 2. Synchronous vs. asynchronous particle updates using 1 and 16 workers with 1particle per worker. Bars denote an average of 10 runs (individual runs denoted left ofthe bars). Depicted is the total runtime for an alignment of the mouse ontology fromthe OAEI anatomy track with itself.
5 Evaluation
Since for large ontologies there are no reference alignments available, effectivenessand scalability of the approach need to be evaluated separately. While it canbe shown for small benchmark ontologies, that the algorithm produces correctresults using a problem specific parameter configuration, a separate experimenthas been conducted evaluating scalability using a cloud infrastructure.
5.1 Effectiveness
Previous evaluations of the MapPSO system at the Ontology Alignment Evalu-ation Initiative (OAEI) campaigns 2008 and 2009 [3, 5] have shown the effective-ness of the approach. In particular with respect to relaxed precision and recallmetrics [7] the OAEI 2009 organisers report that “[. . . ] MapPSO has signifi-cantly better symmetric precision and recall than classical precision and recall,to the point that it is at the level of the best systems.” [9]. The reason for thisdifference to classical metrics is most likely the rather generic parameter config-uration that has been used for all benchmark tests, without adaptation to thesingle alignment scenarios. This generic parameter configuration is a restrictionthat is of minor importance in real-world alignment scenarios.
81
5.2 Scalability
In order to demonstrate the scalability of this approach, an experiment wasconducted aligning two large ontologies from the biomedical domain. The chosenontologies were the Gene Ontology (GO)9 with 31,650 classes, and the MedicalSubject Headings (MESH) Ontology10 with 15,343 classes, both converted tothe OWL format. Specific base matchers were used that take advantage of theclass annotations resulting from the conversion from OBO to OWL.
In the experiment a population of 128 particles was used distributed on 16Amazon EC2 instances (workers) and thus utilising particle pooling by com-puting 8 particles per worker. The EC2 instances were of the type “High-CPUExtra Large Instance, 7GB of memory, 20 EC2 Compute Units (8 virtual coreswith 2.5 EC2 Compute Units each), [. . . ] 64-bit platform, [where] one EC2 Com-pute Unit (ECU) provides the equivalent CPU capacity of a 1.0-1.2GHz 2007Opteron or 2007 Xeon processor”11. The setup was chosen in a way, that thenumber of particles on each instance matches the number of virtual cores, thusenabling efficient multi-threading of particle computation.
The experiment was run for 3.5 h, where due to the asynchronous particleupdates, workers computed between 50 and 75 iterations. The difference in thenumber of iterations illustrates the different computational efforts of each worker.Workers that achieved a lower number of iterations than other workers at thesame time, are most likely computing smaller particles on average. This illus-trates the runtime discrepancy addressed in Sect. 4.1 and the increased parallelefficiency gained by using an asynchronous approach. By using a synchronousapproach the slowest particle constitutes an upper bound and thus a maximumof 50 iterations could have been computed in 3,5 h. The asynchronous approach,however, computes an average of 62.875 iterations, which is an increase of 25%.
Since there is no reference alignment available for GO and MESH as thereis none for any other large ontologies, no statement about the result quality canbe made. The experiment, however, demonstrates the scalability of the proposedapproach, operating on large ontologies and converging to an alignment.
6 Conclusion
The problem of large-scale ontology alignment detection has been addressed bya PSO-based approach. PSO has the characteristics of (i) being independent ofits objective function(s) and thus easily adaptable to various ontology alignmentscenarios, (ii) being incremental, i.e. able to refine an alignment when inputontologies evolve, and (iii) being inherently parallelisable. The latter aspect hasbeen exploited in this paper by deploying the algorithm in the AWS cloud. Thisdeployment is making use of the emerging cloud computing paradigm, which
9 http://www.geneontology.org/10 http://www.nlm.nih.gov/mesh/meshhome.html11 http://aws.amazon.com/ec2/#instance, accessed 2010/06/19
82
provides an on-demand infrastructure for dynamic computational needs, as it isthe case for alignment refinements of gradually evolving ontologies.
When utilising parallel computation infrastructures on a pay-per-use basis,such as cloud offerings, it is crucial to maximise parallel efficiency. Due to thevariable particle sizes used in this approach, different computation times for eachparticle in each iteration can be observed. Thus particle pooling, i.e. multipleparticles per cloud instance, as well as asynchronous particle updates have beenintroduced to the algorithm in order to reduce the time wasted by idling particles.
Previous experiments in the course of OAEI participations have shown theeffectiveness of the PSO-based approach. Complementing these results, the scal-ability via parallelisation has been shown for two large biomedical ontologies.
Having realised the successful deployment of an ontology alignment algorithmin the cloud using an IaaS provider, it is a straightforward extension to provideontology alignment itself as a web service (SaaS) based on a dynamic and scal-able infrastructure. This enables large-scale ontology alignment for every userwithout bothering about hardware requirements. Together with the anytime be-haviour of the PSO-based algorithm, business models comprising the dimensionscomputation time, price, and alignment quality can be created.
Acknowledgement
The presented research was partially funded by the German Federal Ministry ofEconomics (BMWi) under the project Theseus (number 01MQ07019).
References
1. Animoto Productions: animoto - the end of slideshows:. online, http://www.
animoto.com, last seen: 2010/06/162. Bizer, C., Heath, T., Berners-Lee, T.: Linked Data – The Story So Far. Interna-
tional Journal on Semantic Web and Information Systems 5(3), 1–22 (2009)3. Bock, J., Hettenhausen, J.: MapPSO Results for OAEI 2008. In: Proceedings of
the 3rd International Workshop on Ontology Matching (OM-2008). vol. 431. CEURWorkshop Proceedings, http://ceur-ws.org (2008)
4. Bock, J., Hettenhausen, J.: Discrete Particle Swarm Optimisation for OntologyAlignment. Information Sciences (Special Issue on Swarm Intellligence and Appli-cations) (2010), article in press
5. Bock, J., Liu, P., Hettenhausen, J.: MapPSO Results for OAEI 2009. In: Proceed-ings of the 4th International Workshop on Ontology Matching (OM-2009). vol.551, pp. 193–199. CEUR Workshop Proceedings, http://ceur-ws.org (2009)
6. Correa, E.S., Freitas, A.A., Johnson, C.G.: A New Discrete Particle Swarm Algo-rithm Applied to Attribute Selection in a Bioinformatics Data Set. In: Proceedingsof the 8th Genetic and Evolutionary Computation Conference (GECCO-2006). pp.35–42. ACM (2006)
7. Ehrig, M., Euzenat, J.: Relaxed Precision and Recall for Ontology Matching. In:Proceedings of the K-CAP 2005 Workshop on Integrating Ontologies. vol. 156, pp.25–32. CEUR Workshop Proceedings, http://ceur-ws.org (2005)
83
8. Engelbrecht, A.P.: Fundamentals of Computational Swarm Intelligence. Wiley(2007)
9. Euzenat, J., Ferrara, A., Hollink, L., Isaac, A., Joslyn, C., Malaise, V., Meilicke,C., Nikolov, A., Pane, J., Sabou, M., Scharffe, F., Shvaiko, P., Spiliopoulos, V.,Stuckenschmidt, H., Svab-Zamazal, O., Svatek, V., dos Santos, C.T., Vouros, G.,Wang, S.: Results of the Ontology Alignment Evaluation Initiative 2009. In: Pro-ceedings of the 4th International Workshop on Ontology Matching (OM-2009). vol.551. CEUR Workshop Proceedings, http://ceur-ws.org (2009)
10. Euzenat, J., Shvaiko, P.: Ontology Matching. Springer (2007)11. Hilley, D.: Cloud Computing: A Taxonomy of Platform and Infrastructure-level
Offerings. Tech. rep., College of Computing, Georgia Institute of Technology (2009)12. Kennedy, J., Eberhart, R.C.: Particle Swarm Optimization. In: Proceedings of
IEEE International Conference on Neural Networks. vol. 4, pp. 1942–1948. IEEEComputer Society (1995)
13. Kennedy, J., Eberhart, R.C.: Swarm Intelligence. Morgan Kaufmann (2001)14. Kennedy, J., Eberhart, R.C.: A Discrete Binary Version of the Particle Swarm
Algorithm. In: Proceedings of the IEEE International Conference on Systems, Man,and Cybernetics. vol. 5, pp. 4104–4108. IEEE Computer Society (1997)
15. Killalea, T.: Building Scalable Web Services. Queue 6(6), 10–13 (2008)16. Lenk, A., Klems, M., Nimis, J., Tai, S., Sandholm, T.: What’s inside the Cloud? An
architectural map of the Cloud landscape. In: Proceedings of the ICSE Workshopon Software Engineering Challenges of Cloud Computing (CLOUD). pp. 23–31.IEEE Computer Society (2009)
17. Lewis, A., Mostaghim, S., Scriven, I.: Asynchronous Multi-Objective Optimisationin Unreliable Distributed Environments, Studies in Computational Intelligence,vol. 210, pp. 51–78. Springer (2009)
18. Mika, P., Tummarello, G.: Web Semantics in the Clouds. IEEE Intelligent Systems23(5), 82–87 (2008)
19. Newman, A., Li, Y.F., Hunter, J.: Scalable Semantics – the Silver Lining ofCloud Computing. In: Proceedings of the IEEE Fourth International Conferenceon eScience. pp. 111–118. IEEE Computer Society (2008)
20. Qiu, X., Ekanayake, J., Beason, S., Gunarathne, T., Fox, G., Barga, R., Gannon,D.: Cloud Technologies for Bioinformatics Applications. In: Proceedings of the 2ndWorkshop on Many-Task Computing on Grids and Supercomputers (MTAGS). pp.1–10. ACM (2009)
21. Smith, B., Ashburner, M., Rosse, C., Bard, J., Bug, W., Ceusters, W., Goldberg,L.J., Eilbeck, K., Ireland, A., Mungall, C.J., Leontis, N., Rocca-Serra, P., Rutten-berg, A., Sansone, S.A., Scheuermann, R.H., Shah, N., Whetzel, P.L., Lewis, S.:The OBO Foundry: Coordinated Evolution of Ontologies to Support BiomedicalData Integration. Nature Biotechnology 25(11), 1251–1255 (2007)
22. Stein, R., Zacharias, V.: RDF On Cloud Number Nine. In: Proceedings of the 4thWorkshop on New Forms of Reasoning for the Semantic Web: Scalable & Dynamic.pp. 11–23. CEUR Workshop Proceedings, http://ceur-ws.org (2010)
23. Urbani, J., Kotoulas, S., Maassen, J., van Harmelen, F., Bal, H.: OWL Reasoningwith WebPIE: Calculating the Closure of 100 Billion Triples. In: Proceedings of the7th Extended Semantic Web Conference (ESWC). LNCS, vol. 6088, pp. 213–227.Springer (2010)
24. Urbani, J., Kotoulas, S., Oren, E., van Harmelen, F.: Scalable Distributed Rea-soning using MapReduce. In: Proceedings of the 8th International Semantic WebConference (ISWC). LNCS, vol. 5823, pp. 634–649. Springer (2009)
84
Results of theOntology Alignment Evaluation Initiative 2010�
Jerome Euzenat1, Alfio Ferrara2, Christian Meilicke3, Andriy Nikolov4, Juan Pane5,
Francois Scharffe1, Pavel Shvaiko6, Heiner Stuckenschmidt3, Ondrej Svab-Zamazal7,
Vojtech Svatek7, and Cassia Trojahn1
1 INRIA & LIG, Montbonnot, France
{jerome.euzenat,francois.scharffe,cassia.trojahn}@inrialpes.fr2 Universita degli studi di Milano, Italy
[email protected] University of Mannheim, Mannheim, Germany
{christian,heiner}@informatik.uni-mannheim.de4 The Open University, Milton Keynes, UK
[email protected] University of Trento, Povo, Trento, Italy
[email protected] TasLab, Informatica Trentina, Trento, Italy
[email protected] University of Economics, Prague, Czech Republic
{ondrej.zamazal,svatek}@vse.cz
Abstract. Ontology matching consists of finding correspondences between en-
tities of two ontologies. OAEI campaigns aim at comparing ontology matching
systems on precisely defined test cases. Test cases can use ontologies of different
nature (from simple directories to expressive OWL ontologies) and use different
modalities, e.g., blind evaluation, open evaluation, consensus. OAEI-2010 builds
over previous campaigns by having 4 tracks with 6 test cases followed by 15 par-
ticipants. This year, the OAEI campaign introduces a new evaluation modality in
association with the SEALS project. A subset of OAEI test cases is included in
this new modality which provides more automation to the evaluation and more
direct feedback to the participants. This paper is an overall presentation of the
OAEI 2010 campaign.
1 Introduction
The Ontology Alignment Evaluation Initiative1 (OAEI) is a coordinated international
initiative that organizes the evaluation of the increasing number of ontology matching
systems [9]. The main goal of OAEI is to compare systems and algorithms on the same
basis and to allow anyone for drawing conclusions about the best matching strategies.
Our ambition is that from such evaluations, tool developers can improve their systems.
� This paper improves on the “First results” initially published in the on-site proceedings of the
ISWC workshop on Ontology Matching (OM-2010). The only official results of the campaign,
however, are on the OAEI web site.1 http://oaei.ontologymatching.org
85
Two first events were organized in 2004: (i) the Information Interpretation and In-
tegration Conference (I3CON) held at the NIST Performance Metrics for Intelligent
Systems (PerMIS) workshop and (ii) the Ontology Alignment Contest held at the Eval-
uation of Ontology-based Tools (EON) workshop of the annual International Semantic
Web Conference (ISWC) [18]. Then, unique OAEI campaign occurred in 2005 at the
workshop on Integrating Ontologies held in conjunction with the International Confer-
ence on Knowledge Capture (K-Cap) [1]. Starting from 2006 through 2009 the OAEI
campaigns were held at the Ontology Matching workshops collocated with ISWC [8; 7;
3; 6]. Finally in 2010, the OAEI results were presented again at the Ontology Matching
workshop collocated with ISWC, in Shanghai, China2.
The main novelty of this year is the adoption of an environment for automati-
cally processing evaluations (§2.2), which has been developed in coordination with the
SEALS project3. This project aims at providing standardized datasets, a software infras-
tructure for automatically executing evaluations, and evaluation campaigns for typical
semantic web tools, including ontology matching. This year, a subset of OAEI datasets
is included in the SEALS modality. The goal is to provide better direct feedback to the
participants and a more common ground to the evaluation.
We have discontinued the oriented alignment track of the last year because there
was not enough organizational resources to guarantee a satisfying evaluation.
This paper serves as an introduction to the evaluation campaign of 2010 and to
the results provided in the following papers. The remainder of the paper is organized
as follows. In Section 2, we present the overall evaluation methodology that has been
used. Sections 3-7 discuss in turn the settings and the results of each of the test cases.
Section 8 overviews lessons learned from the campaign. Finally, Section 9 outlines
future plans and Section 10 concludes the paper.
2 General methodology
We first present the test cases proposed this year to OAEI participants (§2.1). Then, we
present the evaluation environment, which has been used by participants to test their
systems and launch their evaluation experiments for the campaign (§2.2). Next, we
describe the steps of the OAEI campaign (§2.3-2.6) and report on the general execution
of the campaign (§2.7).
2.1 Tracks and test cases
This year’s campaign has consisted of 4 tracks gathering 6 data sets and different eval-
uation modalities:
The benchmark track (§3): Like in previous campaigns, a systematic benchmark se-
ries have been proposed. The goal of this benchmark series is to identify the areas
in which each alignment algorithm is strong and weak. The test is based on one par-
ticular ontology dedicated to the very narrow domain of bibliography and a number
of alternative ontologies of the same domain for which alignments are provided.
2 http://om2010.ontologymatching.org3 http://www.seals-project.eu
86
The expressive ontologies track offers ontologies using OWL modeling capabilities:
Anatomy (§4): The anatomy real world case is about matching the Adult Mouse
Anatomy (2744 classes) and the NCI Thesaurus (3304 classes) describing the
human anatomy.
Conference (§5): The goal of this track is to find all correct correspondences
within a collection of ontologies describing the domain of organizing con-
ferences (the domain being well understandable for every researcher). Addi-
tionally, ‘interesting correspondences’ are also welcome. Results were evalu-
ated automatically against reference alignments and by data-mining and logical
reasoning techniques. Sample of correspondences and ‘interesting correspon-
dences’ were also evaluated manually.
The directories and thesauri track proposed only web directories this years:
Directory (§6): The directory real world case consists of matching web site di-
rectories (like open directory or Yahoo’s). This year the track consists of two
modalities, the first is composed by more than 4000 elementary tests, and the
second is composed by a single test involving matching of two large directories
(2854 and 6555 nodes each).
Instance matching (§7): The goal of the instance matching track is to evaluate the per-
formance of different tools on the task of matching RDF individuals which originate
from different sources but describe the same real-world entity. Instance matching
is organized in two sub-tasks:
Data interlinking (DI) Participants are requested to re-build the links among the
available RDF resources. Reference alignments are provided for each resource
as RDF alignments.
OWL data track (IIMB & PR): In the OWL data track, data is provided as OWL
individuals according to the RDF/XML format, while reference alignments are
provided as RDF alignments. IIMB is divided into test cases and reference
alignments are automatically generated by introducing controlled modifica-
tions in an initial reference ontology instance. Persons-Restaurants (PR) is a
small real data test case where participants are requested to run matching tools
against two collections of data concerning persons (person1 and person2) and
one collection about restaurants (restaurant1).
The Benchmark, Anatomy and Conference datasets have been evaluated using the
SEALS service. The reason for this is twofold: on the one hand, these data sets are
well known to the organizers and have been used in many evaluations, contrary to the
instance matching data sets, for instance. On the other hand, these data sets come with
a high quality reference alignment which allows for computing the compliance based
measures, such as precision and recall.
This year we had to cancel the VLCR (very large crosslingual resources) task since
it had not enough participants to be retained. The Single task modality in the Directory
track was also canceled due to lack of resources needed to cross check the reference
alignments.
Table 1 summarizes the variation in the results expected from the tests under con-
sideration.
87
test formalism relations confidence modalities language
benchmarks OWL = [0 1] open EN
anatomy OWL = [0 1] open EN
conference OWL-DL =, <= [0 1] blind+open EN
directory OWL = 1 blind+open EN
di RDF = [0 1] open EN
iimb RDF = [0 1] open EN
vlcr SKOS exact-, [0 1] blind DU+EN
+OWL closeMatch expert
Table 1. Characteristics of the test cases (open evaluation is made with already published refer-
ence alignments and blind evaluation is made by organizers from reference alignments unknown
to the participants).
2.2 The SEALS evaluation service
This year, participants have used the SEALS evaluation service for testing their systems
and for launching their own evaluation experiments. A first version of this evaluation
service4 is based on the use of a web service interface wrapping the functionality of a
matching tool to be evaluated [19]. Participants were invited to extend a web service
interface5 and deploy their matchers as web services, which are accessed during the
evaluation process. This setting allows participants for debugging their systems, running
their own evaluations and manipulating the results immediately in a direct feedback
cycle.
In order to start an evaluation, participants had to specify the URL of the matcher
service and the name of the matching system to be evaluated. Then they had to select
the evaluation task to be used (Anatomy, Benchmark or Conference). The specified
web service is validated by the system (two simple ontologies are used to check if the
matcher generates alignments in the correct format). In case of problems, the concrete
validation error is displayed to the user as direct feedback. In case of a successfully
completed validation, the system returns a confirmation message and continues with the
evaluation process. The values of precision, recall and F–measure are then displayed for
each test case.
Furthermore, organizers have a tool for accessing the results registered for the cam-
paign as well as all evaluations being carried out in the evaluation service. Specifically,
results can be visualized and manipulated via an OLAP interface (Figure 1).
2.3 Preparatory phase
Ontologies to be matched and (where applicable) reference alignments have been pro-
vided in advance during the period between June 1st and June 21st, 2010. This gave
potential participants the occasion to send observations, bug corrections, remarks and
other test cases to the organizers. The goal of this preparatory period is to ensure that
the delivered tests make sense to the participants. The final test base was released on
July 8th. The data sets did not evolve after this period.
4 http://seals.inrialpes.fr/platform/5 http://alignapi.gforge.inria.fr/tutorial/tutorial5/
88
Fig. 1. Using OLAP for results visualization.
2.4 Preliminary tests
In this phase, participants were invited to test their systems in order to ensure that the
systems can load the ontologies to be matched and generate the alignment in the correct
format, namely the Alignment format expressed in RDF/XML [5]. Participants have
been requested to provide (preliminary) results by August 30th.
For the SEALS modality, testing could be conducted using the evaluation service
while for the other tracks participants submitted their preliminary results to the orga-
nizers, who analyzed them semi-automatically, often detecting problems related to the
format or to the naming of the required resulting files.
89
2.5 Execution phase
During the execution phase, participants used their systems to automatically match the
ontologies from the test cases. Participants have been asked to use one algorithm and the
same set of parameters for all tests in all tracks. It is fair to select the set of parameters
that provide the best results (for the tests where results are known). Beside parameters,
the input of the algorithms must be the two ontologies to be matched and any general
purpose resource available to everyone, i.e., no resource especially designed for the test.
In particular, participants should not use the data (ontologies and reference alignments)
from other test cases to help their algorithms. In most cases, ontologies are described in
OWL-DL and serialized in the RDF/XML format.
For the standard OAEI modalities, participants had to run their systems on their
own machines and submit the results via mail to the organizers. SEALS participants
ran their systems via the SEALS evaluation service. They obtained a direct feedback on
the results and could validate them as final results. Furthermore, SEALS participants
were invited to register their tools by that time in the SEALS portal6.
Participants also provided the papers that are published hereafter and a link to their
systems and their configuration parameters.
2.6 Evaluation phase
In the evaluation phase, the organizers have evaluated the alignments provided by the
participants and returned comparisons on these results. Final results were due by Oc-
tober 4th, 2010. In the case of blind tests, only the organizers did the evaluation with
regard to the withheld reference alignments.
Concerning SEALS, the participants have used the evaluation service for registering
their results for the campaign. The evaluation effort is minimized due to the fact that
the results are automatically computed by the services in the evaluation service and
organizers have tools for manipulating and visualizing the results.
The standard evaluation measures are precision and recall computed against the
reference alignments. For the matter of aggregation of the measures, we use weighted
harmonic means (weights being the size of the true positives). This clearly helps in the
case of empty alignments. Another technique that has been used is the computation of
precision/recall graphs so it was advised that participants provide their results with a
weight to each correspondence they found. New measures addressing some limitations
of precision and recall have also been used for testing purposes as well as measures for
compensating the lack of complete reference alignments.
2.7 Comments on the execution
Since a few years, the number of participating systems has remained roughly stable: 4
participants in 2004, 7 in 2005, 10 in 2006, 17 in 2007, 13 in 2008, 16 in 2009 and 15
in 2010.
6 http://www.seals-project.eu/join-the-community/
90
The number of covered runs has decreased more than expected: 48 in 2007, 50 in
2008, 53 in 2009, and 37 in 2010. This may be due to the increasing specialization of
tests: some systems are specifically designed for instance matching or for anatomy.
This year many of the systems are validated through web services thanks to the
SEALS evaluation service. For the next OAEI campaign, we expect to be able to ac-
tually run the matchers in a controlled evaluation environment, in order to test their
portability and deployability. This will also allow for comparing system on a same ex-
ecution basis.
The list of participants is summarized in Table 2. Similar to the previous years not
all participants provided results for all tests. They usually did those which are easier to
run, such as benchmark and conference. The variety of tests and the short time given to
provide results have certainly prevented participants from considering more tests.
System Ag
rMak
er
AR
OM
AA
SM
OV
BL
OO
MS
CO
DI
Ef2
Mat
chF
alco
n-A
OG
eRM
eSM
BL
NR
2M
apP
SO
NB
JLM
Ob
ject
Ref
RiM
OM
SO
BO
MT
axo
Map
To
tal=
15
Confidence√ √ √ √ √ √ √ √ √ √ √ √ √ √
benchmarks√ √ √ √ √ √ √ √ √ √ √
11
anatomy√ √ √ √ √ √ √ √ √
9
conference√ √ √ √ √ √ √ √
8
directory√ √ √ √
4
di√ √
2
iimb+pr√ √ √ √ √
5
Total 3 2 5 1 4 3 2 4 1 2 1 1 2 3 3 37
Table 2. Participants and the state of their submissions. Confidence stands for the type of result
returned by a system: it is ticked when the confidence has been measured as non boolean value.
Participants may be divided in two main categories: those who participated in the
instance matching track and those who participated in ontology matching tracks. Three
systems (ASMOV, CODI, RiMOM) participated in both types of tracks. Last year only
two systems (DSSim and RiMOM) had participated in both types of tracks. The sum-
mary of the results track by track is provided in the following sections.
3 Benchmark
The goal of the benchmark data set is to provide a stable and detailed picture of each
algorithm. For that purpose, algorithms are run on systematically generated test cases.
3.1 Test data
The domain of this first test is Bibliographic references. It is based on a subjective view
of what must be a bibliographic ontology. There may be many different classifications
of publications, for example, based on area and quality. The one chosen here is common
91
among scholars and is based on publication categories; as many ontologies (tests #301-
304), it is reminiscent to BibTeX.
The systematic benchmark test set is built around one reference ontology and
many variations of it. The ontologies are described in OWL-DL and serialized in the
RDF/XML format. The reference ontology is that of test #101. It contains 33 named
classes, 24 object properties, 40 data properties, 56 named individuals and 20 anony-
mous individuals. Participants have to match this reference ontology with the variations.
Variations are focused on the characterization of the behavior of the tools rather than
having them compete on real-life problems. They are organized in three groups:
Simple tests (1xx) such as comparing the reference ontology with itself, with another
irrelevant ontology (the wine ontology used in the OWL primer) or the same ontol-
ogy in its restriction to OWL-Lite;
Systematic tests (2xx) obtained by discarding features from some reference ontology.
It aims at evaluating how an algorithm behaves when a particular type of informa-
tion is lacking. The considered features were:
– Name of entities that can be replaced by random strings, synonyms, name with
different conventions, strings in another language than English;
– Comments that can be suppressed or translated in another language;
– Specialization hierarchy that can be suppressed, expanded or flattened;
– Instances that can be suppressed;
– Properties that can be suppressed or having the restrictions on classes dis-
carded;
– Classes that can be expanded, i.e., replaced by several classes or flattened.
Four real-life ontologies of bibliographic references (3xx) found on the web and left
mostly untouched (there were added xmlns and xml:base attributes).
Since one goal of these tests is to offer a permanent benchmark to be used by many,
the test is an extension of the 2004 EON Ontology Alignment Contest, whose test num-
bering is (almost) fully preserved.
The tests are roughly the same as last year. The kind of expected alignments is still
limited: they only match named classes and properties, they mostly use the ”=” relation
with confidence of 1. Full description of these tests can be found on the OAEI web site.
3.2 Results
Eleven systems have participated in the benchmark track of this year’s campaign (see
Table 2). Four systems that had participated last year (AFlood, DSSim, Kosimap and
Lily) did not participate this year, while two new systems (CODI and Ef2Match) have
registered their results.
Table 3 shows the results, by groups of tests. For comparative purposes, the results
of systems that have participated last year are also provided. We display the results of
participants as well as those given by some simple edit distance algorithm on labels
(edna). The full results are on the OAEI web site.
As shown in Table 3, two systems achieve top performances: ASMOV and RiMOM,
with AgrMaker as a close follower, while SOBOM, GeRMeSMB and Ef2Match, re-
92
syst
emre
fali
gn
edn
aA
grM
aker
AR
OM
AA
SM
OV
CO
DI
Ef2
Mat
chF
alco
nG
eRM
eSM
BM
apP
SO
RiM
OM
SO
BO
MT
axo
Map
test
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
Pre
c.R
ec.
1x
x1
.00
1.0
01
.00
1.0
00
.98
1.0
01
.00
0.9
81
.00
1.0
01
.00
0.9
91
.00
1.0
01
.00
1.0
01
.00
1.0
01
.00
1.0
01
.00
1.0
01
.00
1.0
01
.00
0.3
4
2x
x1
.00
1.0
00
.43
0.5
70
.95
0.8
40
.94
0.4
60
.99
0.8
90
.83
0.4
20
.98
0.6
30
.81
0.6
30
.96
0.6
60
.67
0.5
90
.99
0.8
30
.97
0.7
40
.86
0.2
9
3x
x1
.00
1.0
00
.51
0.6
50
.88
0.5
80
.83
0.5
80
.88
0.8
40
.95
0.4
50
.92
0.7
50
.89
0.7
60
.90
0.4
20
.72
0.3
90
.94
0.7
60
.79
0.7
50
.71
0.3
2
H-m
ean
1.0
01
.00
0.4
50
.58
0.9
50
.84
0.9
40
.48
0.9
80
.89
0.8
40
.44
0.9
80
.65
0.8
20
.65
0.9
60
.67
0.6
80
.60
0.9
90
.84
0.9
70
.75
0.8
60
.29
Rel
axed
1.0
01
.00
0.4
50
.58
0.9
50
.84
0.9
40
.48
0.9
90
.89
0.8
40
.44
0.9
80
.65
0.8
20
.65
0.9
60
.67
0.6
80
.60
0.9
90
.84
0.9
70
.75
0.8
60
.29
Wei
gh
ted
1.0
01
.00
0.6
80
.57
0.9
50
.83
0.9
40
.42
0.9
80
.61
0.8
40
.44
0.9
80
.64
0.9
60
.46
0.9
60
.64
0.8
60
.56
0.9
90
.83
0.9
80
.37
0.8
70
.28
Tabl
e3.
Res
ult
so
bta
ined
by
par
tici
pan
tso
nth
eb
ench
mar
kte
stca
se(h
arm
on
icm
ean
s).R
elax
edp
reci
sio
nan
dre
call
corr
esp
on
dto
the
thre
em
easu
res
of
[ 4] :
sym
met
ric
pro
xim
ity,
corr
ecti
on
effo
rtan
do
rien
ted
.T
he
sam
ere
sult
sh
ave
bee
no
bta
ined
usi
ng
thes
eth
ree
mea
sure
s.W
eig
hte
dp
reci
sio
nan
dre
call
takes
into
acco
un
tth
eco
nfi
den
ceas
soci
ated
toco
rres
po
nd
ence
by
the
mat
cher
s.
93
spectively, had presented intermediary values of precision and recall. In the 2009 cam-
paign, Lily and ASMOV had the best results, with aflood and RiMOM as followers,
while GeRoME, AROMA, DSSim and AgrMaker had intermediary performance. The
same group of matchers has been presented in both campaigns. No system had strictly
lower performance than edna.
Looking for each group of tests, in simple tests (1xx) all systems have similar per-
formance, excluding TaxoMap which has presented low value of recall. As noted in
previous campaigns, the algorithms have their best score with the 1xx test series. This
is because there are no modifications in the labels of classes and properties in these
tests and basically all matchers are able to deal with the heterogeneity in labels. For
systematic tests (2xx), which allow better to distinguish the strengths of algorithms,
ASMOV and RiMOM, respectively, achieve the best results, followed by AgrMaker,
SOBOM, GeRMeSMB and Ef2Match, respectively, which have presented good per-
formance, specially in terms of precision. Finally, for real cases (3xx), ASMOV (in
average) provided the best results, with RiMOM and Ef2Match as followers. The best
precision for these cases was obtained by the new participant CODI.
In general, systems have improved their performance since last year: ASMOV
and RiMOM improved their overall performance, AgrMaker and SOBOM signifi-
cantly improved their recall, while MapPSO and GeRMeSBM improved precision.
Only AROMA has significantly decreased in recall. There is no unique set of systems
achieving the best results for all cases, which indicates that systems exploiting different
features of ontologies perform accordingly to the features of each test case.
The results have also been compared with the relaxed measures proposed in [4],
namely symmetric proximity, correction effort and oriented measures (“Relaxed mea-
sures” in Table 3). They are different generalisations of precision and recall in order to
better discriminate systems that slightly miss the target from those which are grossly
wrong. We have used strict versions of these measures (as published in [4] and contrary
to previous years). As Table 3 shows, there is no improvement when comparing classi-
cal and relaxed precision and recall. This can be explained by the fact that participating
algorithms miss the target, by relatively far (the false negative correspondences found
by the matchers are not close to the correspondences in the reference alignment) so the
gain provided by the relaxed measures has no impact.
We have introduced experimentally confidence-weighted precision and recall in
which correspondences are weighted by the confidence matchers put on it. If the confi-
dence is 1., then the correspondence scores exactly like in classical precision and recall.
Otherwise, it scores for the amount of confidence. If the correspondence is correct, this
will contribute to decrease recall – it will be counted for less than 1. –, if the correspon-
dence is incorrect, this will increase precision – by counting the mistake for less than 1.
So this rewards systems able to provide accurate confidence measures (or penalizes less
mistakes on correspondences with low confidence). These measures provide precision
increase for Falcon, MaPSO and edit distance (which had apparently many incorrect
correspondences with low confidence), and recall decrease for Falcon, ASMOV and
SOBOM (which had apparently many correct correspondences with low confidence).
There are only little variation for other systems. As expected, CODI, which confidence
was always 1, shows no variation.
94
As last year, many algorithms provided their results with confidence measures. It
is thus possible to draw precision/recall graphs in order to compare them. Figure 2
shows the precision and recall graphs of this year. These results are only relevant for
the results of participants who provide confidence measures different from 1 or 0 (see
Table 2). Contrary to previous years these graphs are not drawn with the same principles
as TREC’s. They now show the real precision at n% recall and they stop when no more
correspondences are available (then the end point corresponds to the precision and recall
reported in Table 3). The values are not anymore an average but a real precision and
recall over all the tests. The numbers in the legend are the Mean Average Precision
(MAP): the average precision for each correct retrieved correspondence. These new
graphs represent well the effort made by the participants to keep a high precision in
their results, and to authorize a loss of precision with a few correspondences with low
confidence.
The results presented in Table 3 and those displayed in Figure 2 single out the same
group of systems, ASMOV, RiMOM and AgrMaker, which perform these tests at the
highest level. Out of these, ASMOV has slightly better results than the two others. So,
this confirms the previous observations on raw results.
95
recall0. 1.0.
precision
1.
refalign
1.00
edna
0.58
AgrMaker
0.84
AROMA
0.47
ASMOV
0.88
CODI
0.44
Ef2Match
0.65
Falcon
0.63
GeRMeSMB
0.66MapPSO
0.59
RiMOM
0.84
SOBOM
0.74TaxoMap
0.29
Fig. 2. Precision/recall graphs for benchmarks. The results given by the participants are cut under
a threshold necessary for achieving n% recall and the corresponding precision is computed. Sys-
tems for which these graphs are not meaningful (because they did not provide graded confidence
values) are drawn in dashed lines.
96
4 Anatomy
The anatomy track confronts matching technology with a specific type of ontologies
from the biomedical domain. In this domain, a significant number of ontologies have
been built covering different aspects of medical research.
4.1 Test data and experimental setting
The data set of this track has been used since 2007. For a detailed description we re-
fer the reader to the OAEI 2007 [7] results paper. As in previous years, we divided
the matching task into four subtasks. Subtask #1 is compulsory for participants of the
anatomy track, while subtask #2, #3 and #4 are again optional tasks.
Subtask #1 The matcher has to be applied with its standard settings.
Subtask #2 An alignment has to be generated that favors precision over recall.
Subtask #3 An alignment has to be generated that favors recall over precision.
Subtask #4 A partial reference alignment has to be used as additional input.
Notice that in 2010 we used the SEALS evaluation service for subtask #1. In the course
of using the SEALS services, we published the complete reference alignment for the
first time. In future, we plan to include all subtasks in the SEALS modality. This re-
quires to extend the interfaces of the SEALS evaluation service to allow for example an
(incomplete) alignment as additional input parameter.
The harmonization of the ontologies applied in the process of generating a reference
alignment (see [2] and [7]) resulted in a high number of rather trivial correspondences
(61%). These correspondences can be found by very simple string comparison tech-
niques. At the same time, we have a good share of non-trivial correspondences (39%).
This is an important characteristic of the data set to be taken into account in the fol-
lowing analysis. The partial reference alignment used in subtask #4 is the union of all
trivial correspondences and 54 non-trivial correspondences.
We slightly improved the test data set for the 2010 evaluation. We removed some
doubtful subsumption axioms and added a number of disjointness statements at the top
of the hierarchies to increase the expressivity of the data set. Furthermore, we eliminated
some incorrect correspondences.7
In previous years, we reported about runtimes measured by the participants. The
differences we observed – from several minutes to several days – could not be explained
by the use of different hardware. However, these differences became less significant
over the years. Therefore, we abstained from an analysis of runtimes this year. In 2011,
we plan to execute the matching systems on the SEALS platform to enable an exact
measurement of runtimes not biased by differences in hardware equipment. So far we
refer the reader interested in runtimes to the result papers of the participants.
7 We gratefully thank Elena Beisswanger (Jena University Language and Information Engineer-
ing Lab) for her thorough support on improving the quality of the data set. The modifications
are documented at http://webrum.uni-mannheim.de/math/lski/anatomy10/modifications2010.html
97
4.2 Results
While the number of participants has been roughly stable over four years, we had in
2010 more systems that participated for the first time (5 systems) than in previous years
(in average 2 systems; see Table 4 for an overview). Three of the newcomers participate
also in other tracks, while NBJLM and BLOOMS participate only in the Anatomy track.
Notice also that AgreementMaker (AgrMaker) uses a track-specific parameter setting.
Taking part in several tracks with a standard setting makes it obviously much harder to
obtain good results in a specific track.
System 2007 2008 2009 2010AFlood
√ √AgrMaker
√+ +
AROMA√ √
AOAS +
ASMOV√ √ √ √
BLOOMS +
CODI√
DSSim√ √ √
Ef2Match +
Falcon AO√
GeRMeSMB√
Kosimap√
Lily√ √ √
NBJLM +
Prior+√
RiMOM√
+√
SAMBO + +
SOBOM + +
TaxoMap√ √ √
+
X SOM√
Avg. F-measure 0.598 0.718 0.764 0.785
Table 4. Overview on anatomy participants from 2007 to 2010, a√
-symbol indicates that the
system participated, + indicates that the system achieved an F-measure ≥ 0.8 in subtask #1.
In the last row of Table 4, the average of F-measures per year in subtask #1 is shown.
We observe significant improvements over time. However, the measured improvements
decrease over time and seem to reach a top (2007 +12% → 2008 +5% → 2009 +2%
→ 2010). We have marked the participants with an F-measure ≥ 0.8 with a + symbol.
Note that in each of the previous years, only two systems reached this level, while in
2010 six systems reached a higher value than 0.8.
Main results for subtask #1. The results for subtask #1 are presented in Table 5
ordered with respect to the achieved F-measure. In 2010, AgreementMaker (AgrMaker)
generated the best alignment with respect to F-measure. This result is based on a high
recall compared to the systems on the following positions. Even the SAMBO system of
2007 could not generate a higher recall with the use of UMLS.
98
Task #1 Task #2 Task #3 Recall+System Prec. F Rec. Prec. F Rec. Prec. F Rec. #1 #3AgrMaker* 0.903 0.877 0.853 0.962 0.843 0.751 0.771 0.819 0.874 0.630 0.700
Ef2Match 0.955 0.859 0.781 0.968 0.842 0.745 0.954 0.859 0.781 0.440 0.440
NBJLM* 0.920 0.858 0.803 - - - - - - 0.569 -
SOBOM 0.949 0.855 0.778 - - - - - - 0.433 -
BLOOMS 0.954 0.828 0.731 0.967 0.829 0.725 - - - 0.315 -
TaxoMap 0.924 0.824 0.743 0.956 0.801 0.689 0.833 0.802 0.774 0.336 0.414
ASMOV 0.799 0.785 0.772 0.865 0.808 0.757 0.717 0.753 0.792 0.470 0.538
CODI 0.968 0.779 0.651 0.964 0.785 0.662 0.782 0.736 0.695 0.182 0.383
GeRMeSMB 0.884 0.456 0.307 0.883 0.456 0.307 0.080 0.147 0.891 0.249 0.838
Table 5. Results for subtasks #1, #2 and #3 in terms of precision, F-measure, and recall (in
addition recall+ for #1 and #3). Systems marked with a * do not participate in other tracks or
have chosen a setting specific to this track. Note that ASMOV modified its standard setting in a
very restricted way (activating UMLS as additional resource). Thus, we did not mark this system.
AgreementMaker is followed by three participants (Ef2Match, NBJLM and
SOBOM) that clearly favor precision over recall. Notice that these systems obtained
better scores or scores that are similar to the results of the top systems in the previous
years. One explanation can be seen in the fact that the organizers of the track made the
reference alignment available to the participants. More precisely, participants could at
any time compute precision and recall scores via the SEALS services to test different
settings of their algorithms. This allows to improve a matching system in a direct feed-
back cycle, however, it might happen that a perfect configuration results in problems
for different data sets.
Recall+ and further results. We use again the recall+ measure as defined in [7]. It
measures how many non trivial correct correspondences, not detectable by string equiv-
alence, can be found in an alignment. The top three systems with respect to recall+
regarding subtask #1 are AgreementMaker, NBJLM and ASMOV. Only ASMOV has
participated in several tracks with the same setting. Obviously, it is not easy to find a
large amount of non-trivial correspondences with a standard setting.
In 2010, six systems participated in subtask #3. The top three systems regarding
recall+ in this task are GeRoMe-SMB (GeRMeSMB), AgreementMaker and ASMOV.
Since a specific instruction about the balance between precision and recall is missing in
the description of the task, the results vary to a large degree. GeRoMe-SMB detected
83.8% of the correspondences marked as non trivial, but at a precision of 8%. Agree-
mentMaker and ASMOV modified their settings only slightly, however, they were still
able to detect 70% and 53.8% of all non trivial correspondences.
In subtask #2, seven systems participated. It is interesting to see that systems like
ASMOV, BLOOMS and CODI generate alignments with slightly higher F-measure for
this task compared to the submission for subtask #1. The results for subtask #2 for
AgreementMaker are similar to the results submitted by other participants for subtask
#1. This shows that many systems in 2010 focused on a similar strategy that exploits
the specifics of the data set resulting in a high F-measure based on a high precision.
99
Subtask #4. In the following, we refer to an alignment generated for subtask #n as
An. In our evaluation we use again the method introduced in 2009. We compare both
A1 ∪Rp and A4 ∪Rp with the reference alignment R.8 Thus, we compare the situation
where the partial reference alignment is added after the matching process against the
situation where the partial reference alignment is available as additional resource ex-
ploited within the matching process. Note that a direct comparison of A1 and A4 would
not take into account in how far the partial reference alignment was already included in
A1 resulting in a distorted interpretation.
System Δ-Precision Δ-F-measure Δ-RecallAgrMaker +0.025 0.904→0.929 −0.002 0.890→0.888 −0.025 0.876→0.851
ASMOV +0.029 0.808→0.837 +0.006 0.816→0.822 −0.016 0.824→0.808
CODI −0.002 0.970→0.968 +0.019 0.824→0.843 +0.030 0.716→0.746
SAMBOdtf2008 +0.021 0.837→0.856 +0.011 0.852→0.863 +0.003 0.867→0.870
Table 6. Changes in precision, F-measure and recall based on comparing A1∪Rp and A4 against
reference alignment R.
Results are presented in Table 6. Three systems participated in task #4 in 2010.
Additionally, we added a row for the 2008 submission of SAMBOdtf. This system
had the best results measured in the last years. AgreementMaker and ASMOV use the
input alignment to increase the precision of the final result. At the same time these sys-
tems filter out some correct correspondences, finally resulting in a slightly increased
F-measure. This fits with the trend observed in the past years (compare with the results
for SAMBOdtf in 2008). The effects of this strategy are not very strong. However, as
argued in previous years, the input alignment has a characteristics that makes hard to
exploit this information. CODI has chosen a different strategy. While changes in pre-
cision are negligible, recall increases by 3%. Even though the overall effect is still not
very strong, the system exploits the input alignment in the most effective way. However,
the recall of CODI for subtask #1 is relatively low compared to the other systems. It is
unclear whether the strategy of CODI would also work for the other systems where a
ceiling effect might prevent the exploitation of the positive effects.
4.3 Conclusions
Overall, we see a clear improvement comparing 2010 results with the results of previous
years. This holds both for the “average participant” as well as for the top performer. A
very positive outcome can be seen in the increased recall values. In addition to the
evaluation experiments reported so far, we computed the union of all submissions to
subtask #1. For the resulting alignment we measured a precision of 69.7% and a recall
of 92.7%. We added additionally the correct correspondences generated in subtask #3
and reached a a recall of 97.1%. Combining the strategies used by different matching
systems it is thus possible to detect nearly all correct correspondences.
8 We use A4 ∪ Rp – instead of using A4 directly – to ensure that a system, which does not
include the input alignment in the output, is not penalized.
100
5 Conference
The conference test set introduces matching several moderately expressive ontologies.
Within this track, participant results were evaluated using diverse evaluation methods.
The evaluation has been supported by the SEALS evaluation service this year.
5.1 Test data
The data set of this track has been extended with one ontology being in OWL 2.9 For a
data set description we refer the reader to the OAEI 2009 results paper [6].
5.2 Results
We had eight participants: AgreementMaker (AgrMaker), AROMA, ASMOV, CODI,
Ef2Match, Falcon, GeRMeSMB and SOBOM. All participants delivered all 120 align-
ments. CODI delivered ‘certain’ correspondences, the other matchers delivered corre-
spondences with graded confidence values between 0 and 1.
Evaluation based on the reference alignments. We evaluated the results of partici-
pants against reference alignments. They include all pairwise combinations between 7
different ontologies, i.e. 21 alignments.
matcher confidence threshold Prec. FMeas. Rec.
AgrMaker 0.66 .53 .58 .62
AROMA 0.49 .36 .42 .49
ASMOV 0.22 .57 .60 .63
CODI * .86 .62 .48
Ef2Match 0.84 .61 .60 .58
Falcon 0.87 .74 .59 .49
GeRMeSMB 0.87 .37 .43 .51
SOBOM 0.35 .56 .56 .56
Table 7. Confidence threshold, precision and recall for optimal F-measure for each matcher.
For a better comparison, we established the confidence threshold which provides
the highest average F-measure (Table 7). Precision, recall, and F-measure are given
for this optimal confidence threshold. The dependency of F-measure on the confidence
threshold can be seen from Figure 3. There is one asterisk in the column of confidence
threshold for matcher CODI which did not provide graded confidence.
In conclusion, the matcher with the highest average F-measure (62%) is CODI
which did not provide graded confidence values. Other matchers are very close to this
score (e.g. ASMOV with F-Measure 0.60, Ef2Match with F-Measure 0.60, Falcon with
F-Measure 0.59). However, we should take into account that this evaluation has been
made over a subset of all alignments (one fifth).
9 Ontologies have been developed within the OntoFarm project http://nb.vse.cz/
˜svatek/ontofarm.html
101
Fig. 3. F-measures depending on confidence.
Comparison with previous years. Three matchers also participated in the last two years.
ASMOV participated in all three consecutive years with increasing highest average F-
measure: from .43 in 2008 and .47 in 2009 to .60 in 2010. AgreementMaker participated
with .57 in 2009 and with .58 in 2010 regarding highest average F-measure. Finally,
AROMA participated with the same highest average F-measure in both 2009 and 2010.
Evaluation based on posterior manual labeling. This year we took the most se-
cure correct correspondences, i.e., with the highest confidence, as a population for each
matcher. Per matcher, we evaluated 100 correspondences randomly chosen from all cor-
respondences of all 120 alignments with confidence 1.0 (sampling). Because AROMA,
ASMOV, Falcon, GeRMeSMB and SOBOM do not have enough correspondences with
1.0 confidence we took the 100 correspondences with highest confidence. For all of
these matchers (except ASMOV where we found exactly 100 correspondences with
highest confidence values) we sampled over their population.
Table 8 presents approximated precisions for each matcher over its population of
best correspondences. N is the population of all best correspondences for one matcher.
n is a number, ideally 100, of randomly chosen correspondences, among the best cor-
respondences for each matcher. TP is a number of correct correspondences from the
sample, and P* is an approximation of precision for the correspondences in each popu-
lation; additionally there is a margin of error computed as:
√(N/n)−1√
Nbased on [20].
102
matcher AgrMaker AROMA ASMOV CODI Ef2Match Falcon GeRMeSMB SOBOM
N 804 108 100 783 1236 127 110 105
n 100 100 100 100 100 100 100 100
TP 92 68 86 98 79 96 30 82
P* .92 .68 .86 .98 .79 .96 .30 .82±9.4 ±2.7 ±9.3 ±9.6 ±4.6 ±3.0 ±2.2
Table 8. Approximated precision for 100 best correspondences for each matcher.
From Table 8 we can conclude that CODI, Falcon and AgreementMaker have the
best precision (higher than 90%) over their 100 more confident correspondences.
Evaluation based on data mining supported with matching patterns. Results of
OAEI participants, i.e. correspondences, contain several attributes characterizing these
correspondences from different aspects. Additionally, there is also information in which
matching patterns the given correspondence participates; for details about this see [15;
22].
In total there are nine matching patterns (MP1 - MP9). MP4, MP5, and MP6are inspired by correspondence patterns from [16]. In principle, it is not possible to
say which matching pattern is desirable or not desirable. This must be decided on the
basis of an application context or possible alternatives. However, we could roughly say
that while MP2 and MP5 seems to be desirable, MP7, MP8, and MP9 indicate
incorrect correspondences related to inconsistency (see section below).
Antecedent Succedent Values
System Certainty factor Resource1 Resource2 Result Supp AvgDff
t1 AgrMaker < 1.0; 1.0 > * * + 0.024 0.95
t2 ASMOV < 0.4; 0.8) * * + 0.01 0.6
t3 SOBOM < 0.4 * * + 0.024 0.37
t4 ASMOV < 0.4 * w - 0.013 1.25
t5 SOBOM * * w - 0.014 1.21
Table 9. Hypotheses for tasks 1 and 2.
Antecedent Succedent Values
System ResultMP Supp AvgDff
m1 ASMOV MP2 0.013 0.79
m2 AgrMaker MP2 0.015 0.1
m3 AROMA MP4 0.016 0.16
Table 10. Association hypotheses related to matching patterns.
For data mining we employed the 4ft-Miner procedure of the LISp-Miner data min-
ing system10 for mining of association rules. We found several interesting association
10 http://lispminer.vse.cz/
103
hypotheses: t1 to t5 are related to confidence value or underlying resources of ontolo-
gies (see Table 9) and m1 to m3 are related to matching patterns (see Table 10). In
total, there were 16522 correspondences in the data matrix. For instance we can inter-
pret hypothesis t1 as follows correspondences that are produced by AgrMaker and havethe highest confidence value (i.e. 1.0) are by 95% (i.e. almost twice) more often correctthan correspondences produced by all systems with all confidence values (on average).
In conclusion, regarding first three hypotheses we could say that AgrMaker is quite
sure about correspondences with the highest value than other matchers. On the other
side, ASMOV is surprisingly correct about correspondences with lower confidence val-
ues than other matchers. These hypotheses confirmed findings from the previous year
since these two matchers also participated in the OAEI-2009. SOBOM is more correct
for correspondences with the lowest confidence values. Hypotheses t4 and t5 point out
that ASMOV and SOBOM work worse with ontologies based on the web. Regarding
hypotheses containing matching patterns, ASMOV and AgrMaker found MP2, while
AROMA found MP4 more often than other systems. In comparison with the previous
year, there are no interesting hypotheses with matching patterns related to inconsis-
tency. This can be explained by the fact that many occurrences of matching patterns
related with inconsistency is generally lower than in the previous year of OAEI. These
findings roughly correspond with the results of the evaluation based on the alignment
coherence in the next section.
Evaluation based on alignment coherence. In the following we apply the Maximum
Cardinality measure as proposed in [14] to measure the degree of alignment incoher-
ence. Results are depicted in Table 11 which shows the average for all testcases of the
conference track except the testcases where the ontologies confious and linklings are
involved. These ontologies resulted in some combinations of ontologies and alignments
in reasoning problems. Note that we did not use the original alignments, but the align-
ments with optimal threshold. However, the average size of the resulting alignment still
varies to a large degree.
Matcher AgrMaker AROMA ASMOV CODI Ef2Match Falcon GeRMeSMB SOBOM
Max-Card % >14.8% >17.5% 5.6% 0.1% 7.2% >4.8% >12.6% >10.7
N 17.1 16.4 18.2 6.9 12.8 8.9 18.2 11.7
Table 11. Degree of incoherence and size of alignment in average for the optimal a posteriori
threshold. The prefix > is added whenever the search algorithm stopped in one of the testcase
due to a timeout of 1000 seconds prior to finding the solution.
Compared to the other participants CODI generates the lowest degree of incoher-
ence. This result is partially caused by the small size of alignments that make the oc-
currence of an incoherence less probable. Taking this into account, the ASMOV system
achieves a remarkable result. Even though the alignments of ASMOV comprise the
highest number of correspondences, the degree of incoherence 5.6% is relatively small
due to the verification component built into the system [12]. Overall it is a suprising
result that still only few matching systems take alignment incoherence into account.
104
6 Directory
The directory test case aims at providing a challenging task for ontology matchers in the
domain of large directories to show whether ontology matching tools can effectively be
applied for the integration of “shallow ontologies”. This task focusses on evaluating the
performances of existing matching tools in a real world taxonomy integration scenario.
6.1 Test set
As in previous years, the data set exploited in the directory matching task was con-
structed from the Google, Yahoo and Looksmart web directories following the method-
ology described in [10]. The data set is presented as taxonomies where the nodes of the
web directories are modeled as classes and classification relation connecting the nodes
is modeled as rdfs:subClassOf.
The key idea of the data set construction methodology is to significantly reduce the
search space for human annotators. Instead of considering the full matching task which
is very large (Google and Yahoo directories have up to 3 ∗ 105 nodes each: this means
that the human annotators need to consider up to (3∗105)2 = 9∗1010 correspondences),
it uses semi automatic pruning techniques in order to significantly reduce the search
space. For example, for the data set described in [10], human annotators consider only
2265 correspondences instead of the full matching problem.
The specific characteristics of the data set are:
– Simple relationships. Basically web directories contain only one type of relation-
ship called “classification relation”.
– Vague terminology and modeling principles: The matching tasks incorporate the
typical “real world” modeling and terminological errors.
– More than 4.500 node matching tasks, where each node matching task is composed
of the paths to root of nodes in the web directories.
– Reference correspondences for the equivalence relation for all the matching tasks.
6.2 Results
In OAEI-2010, 3 out of 15 matching systems participated in the web directories test
case, while in OAEI-2009 7 out of 16, in OAEI-2008, 7 out of 13, in OAEI-2007, 9 out
of 17, in OAEI-2006, 7 out of 10, and in OAEI-2005, 7 out of 7 did it. The systems that
submitted their results to the Single task modality of the Directory track were ASMOV,
GeRoMe-SMB, MapPSO and TaxoMap, though the task was canceled due to lack of
resources needed to cross check the reference alignments.
Precision, F-measure and recall results of the systems are shown in Figure 4. These
indicators have been computed following the TaxMe2 [10] methodology, with the help
of the Alignment API [5], version 4.0.
We can observe that ASMOV has maintained its recall, but increased its precision
by 1 point in comparison to 2009. MapPSO has increased its recall (+27) and precision
(+7) values, resulting in a 20 points increase in the F-measure from its last participation
in 2008. TaxoMap has decreased its recall (-7) but increased its precision (+3), resulting
105
Fig. 4. Matching quality results.
in an overall decrease of F-measure (-6) from its last participation in 2009. ASMOV is
the system with the highest F-measure value in 2010.
In total, 24 matching systems have participated during the 6 years (2005 - 2010)
of the OAEI campaign in the directory track where 40 individual submissions from
different systems have been received over the past 6 years. No single system has par-
ticipated in all campaigns involving the web directory dataset (2005 - 2010). A total of
15 systems have participated only one time in the evaluation, 5 systems have partici-
pated 3 times (DSSIM, Falcon, Lily, RiMOM and TaxoMap), and only 1 system has
participated 4 times (ASMOV).
As can be seen in Figure 5, this year there is an small increase (2%) in the average
precision, in comparison to 2007 and 2008. The average recall in 2010 increased in
comparison to 2009, reaching the same highest average recall value as in 2007. Con-
sidering F-measure, results for 2010 show the highest average in the 5 years (2006 to
2010). Notice that in 2005 the data set allowed only the estimation of recall, therefore
Figure 5 does not contain values of precision and F-measure for 2005.
Fig. 5. Average results of the participating systems per year.
106
A comparison of the results from 2006 - 2010 for the top-3 systems of each year
based on the highest values of the F-measure indicator is shown in Figure 6. An impor-
tant note is that since there are only 3 participants this year, they all made their ways into
the top three. The comparison of the top three participants has being made since 2006,
therefore we keep the same comparison (and not the top 2, for example) for historical
reasons. The quality of the best F-measure result of 2010 (0.63) achieved by ASMOV is
equal to the best F-measure of 2009 by the same system, higher than the best F-measure
of 2007 by DSSim (0.49) and than that of 2006 by Falcon (0.43), but still lower than the
best F-measure of 2007 (0.71) by OLA2. All three participating systems have achieved
the same precision in 2010 (0.61), but this precision is lower than the best values of
2009 (0.62) by kosimap, in 2008 (0.64) by ASMOV and in 2007 by both OLA2 and
X-SOM. Finally, for what concerns recall, the best result of 2010 achieved by ASMOV
(0.65) is equal to the best value of 2009 (0.65) also achieved by ASMOV, higher than
the best value of 2008 (0.41) demonstrated by DSSim and the best value in 2006 (0.45)
by Falcon, but still lower than the best result obtained in 2007(0.84) by OLA2.
Fig. 6. Comparison of matching quality results in 2006 - 2010.
Partitions of positive and negative correspondences, according to the system results,
are presented in Figure 7 a) and Figure 7 b), respectively. Figure 7 a) shows that the
systems managed to discover only 67% of the total number of positive correspondences
(Nobody = 33%). Only 27% of positive correspondences were found by all three par-
ticipating systems. The percentage of positive correspondences found by the systems
this year is slightly lower than the values of 2009, when 68% of the positive corre-
spondences where found [6], but still higher than the values of 2008, when 54% of the
positive correspondences where found [3].
Figure 7 b) shows that more than half (59%) of the negatives correspondences were
not found by the systems (correctly) in comparison to 56% not found in 2009). Fig-
ure 7 b) also shows that all participating systems found 16% of the negative correspon-
dences, i.e., mistakenly returned them as positive, in comparison to 17% in 2009. These
two observations explain the small increase in precision in Figure 5. The last two ob-
servations also suggest that the discrimination ability of the dataset remains as high as
in previous years.
107
a) b)Fig. 7. Partition of the system results: a) on positive correspondences, b) on negative correspon-
dences.
Figure 7 a) shows that 33% of positive correspondences have not been found by any
of the matching systems this year. This value is better that the values of 2006 (43%) and
2008 (46%) but worse than of 2009 (32%). In 2007, all the positive correspondences
have been collectively found; these results (2007) were exceptional because the par-
ticipating systems altogether had a full coverage of the expected results and very high
precision and recall. Unfortunately, the best systems of 2007 did not participate in the
last years and the other systems do not seem to cope with the results of 2007.
Figure 7 b) shows that this year 59% of the negatives correspondences were cor-
rectly not found. There is an increase in comparison to the value of 2009 (56%) but
a decrease in comparison to the value of 2008, when 66% of the negatives correspon-
dences where not found, being the best value in all years (2006 to 2010). This year
16% of the negative correspondences were mistakenly found by all the (3) participat-
ing systems, being the best value that of 2008 (1% for all (7) participating systems).
An interpretation of these observations could be that the set of participating systems in
2010 seems to have found a good balance between being “cautious” (not finding nega-
tives) and being “brave” (finding positives), resulting in average increases on precision,
recall and F-measure as shown in Figure 5. In average, in 2010 the participants have a
more “cautious” strategy of all years except 2008, being a little bit more “brave” than
in 2007 and 2008. In 2007, we can observe that the set of systems showed the most
“brave” strategy in discovering correspondences of all the yearly evaluation initiatives,
when the set of positive correspondences was fully covered, but covering also 98% of
the negative correspondences.
6.3 Comments
This year the average performance of the participants (given by the increase in preci-
sion and F-measure in Figure 5) is the best of all 5 years (2006 to 2010). This suggests
that the set of participating systems has found a balance between a “brave and cau-
tious” behavior for discovering correspondences. However, the value for the F-measure
(0.53) indicates that there is still room for improvements. In comparison to 2009, there
is an increase of 2% in F-measure where the average F-measure was (0.51). Finally, as
partitions of positive and negative correspondences indicate (see Figure 7 a) and Fig-
ure 7 b)), the dataset still retains a good discrimination ability, i.e., different sets of
correspondences are still hard for the different systems.
108
7 Instance matching
The instance matching track was included into the OAEI campaigns for the second
time. The goal of the track is to evaluate the performance of different tools on the task
of matching RDF individuals which originate from different sources but describe the
same real-world entity. With the development of the Linked Data initiative, the growing
amount of semantic data published on the Web and the need to discover identity links
between instances from different repositories, this problem gained more importance in
the recent years. Unlike the other tracks, the instance matching tests specifically focus
on an ontology ABox. However, the problems which have to be resolved in order to
match instances correctly can originate at the schema level (use of different properties
and classification schemas) as well as at the data level, e.g., different format of values.
This year, the track included two tasks. The first task (data interlinking - DI) aims at
testing the performance of tools on large-scale real-world datasets published according
to the Linked Data principles. The second one (IIMB & PR) uses a set of artificially
generated and real test cases respectively. These are designed to illustrate all common
cases of discrepancies between individual descriptions (different value formats, modi-
fied properties, different classification schemas). The list of participants to the Instance
Matching track is shown in Table 12.
System DI IIMB SMALL IIMB LARGE PRASMOV
√ √ √ASMOV D
√CODI
√ √ √LN2R
√ObjectCoref
√ √RiMOM
√ √ √ √
Table 12. Participants in the instance matching track.
7.1 Data interlinking task (DI)
Data interlinking is known under many names according to various research communi-
ties: equivalence mining, record linkage, object consolidation and coreference resolu-
tion to mention the most used ones. In each case, these terms are used for the task of
finding equivalent entities in or across datasets. As the quantity of datasets published on
the Web of data dramatically increases, the need for tools helping to interlink resources
becomes more critical. It is particularly important to maximize the automation of the
interlinking process in order to be able to follow this expansion.
This year, we propose to interlink four datasets together. We have selected datasets
for their potential to be interlinked, for the availability of curated interlinks between
them, and for their size. All datasets are on the health-care domain and all of them
contain information about drugs (see [13] for more details on the datasets):
dailymed is published by the US National Library of Medecine and contains informa-
tion about marketed drugs. Dailymed contains information on the chemical struc-
ture, mechanism of action, indication, usage, contraindications and adverse reac-
tions for the drugs.
109
!"!!# !"$!# !"%!# !"&!# !"’!# !"(!# !")!# !"*!# !"+!# !",!# $"!!#
-./012134#
./0566#
789 /52:./#
-./012134#
./0566#
789 /52:./#
;<=/0>?3./@#
A1B;B#
C89 /54# 21D/.# D.:E<54F# D12/5239 /# D516G9 /D#
Fig. 8. Results of the DI task.
diseasome contains information about 4300 disorders and genes.
drugbank is a repository of more than 5000 drugs approved by the US Federal Drugs
Agency. It contains information about chemical, pharmaceutical and pharmacolog-
ical data along with the drugs data.
sider contains information on marketed drugs and their recorded adverse reactions. It
was originally published on flat files before being converted as linked-data through
a relational database.
These datasets were semi-automatically interlinked using Silk [21] and ODD Linker[11] providing the reference alignments for this task and participants were asked to
retrieve these links using an automatic method.
Only two systems participated in the data interlinking task, probably due to the
difficulties of matching large collections of data: ObjectCoref and RiMOM. The results
of these systems are shown in Figure 8.
The results are very different for two systems, with ObjectCoref being better in pre-
cision and RiMOM being better in recall. A difficult task with real interlinked data is to
understand if the results are due to a weakness of the matching system or because links
can be not very reliable. In any case, what we can conclude from this experiment with
linked data is that a lot of work is still required in three directions: i) providing a reli-
able mechanism for systems’ evaluation; ii) improving the performances of matching
systems in terms of both precision and recall; iii) work on the scalability of matching
techniques in order to make affordable the task of matching large collections of real
data. Starting from these challenges, data interlinking will be one one the most impor-
tant future directions for the instance matching evaluation initiative.
110
7.2 OWL data task (IIMB & PR)
The OWL data task is focused on two main goals:
1. to provide an evaluation dataset for various kinds of data transformations, including
value transformations, structural transformations and logical transformations;
2. to cover a wide spectrum of possible techniques and tools.
To this end, we provided two groups of datasets, the ISLab Instance Matching
Benchmark (IIMB) and the Person-Restaurants benchmark (PR). In both cases, par-
ticipants were requested to find the correct correspondences among individuals of the
first knowledge base and individuals of the other. An important task here is that some
of the transformations require automatic reasoning for finding the expected alignments.
IIMB. IIMB is composed of a set of test cases, each one represented by a set of in-
stances, i.e., an OWL ABox, built from an initial dataset of real linked data extracted
from the web. Then, the ABox is automatically modified in several ways by generating
a set of new ABoxes, called test cases. Each test case is produced by transforming the
individual descriptions in the reference ABox in new individual descriptions that are
inserted in the test case at hand. The goal of transforming the original individuals is
twofold: on one side, we provide a simulated situation where data referring to the same
objects are provided in different data sources; on the other side, we generate different
datasets with a variable level of data quality and complexity. IIMB provides transfor-
mation techniques supporting modifications of data property values, modifications of
number and type of properties used for the individual description, and modifications
of the individuals classification. The first kind of transformations is called data valuetransformation and it aims at simulating the fact that data expressing the same real
object in different data sources may be different because of data errors or because of
the usage of different conventional patterns for data representation. The second kind
of transformation is called data structure transformation and it aims at simulating the
fact that the same real object may be described using different properties/attributes in
different data sources. Finally, the third kind of transformation, called data semantictransformation, simulates the fact that the same real object may be classified in differ-
ent ways in different data sources.
The 2010 edition of IIMB is a collection of OWL ontologies consisting of 29 con-
cepts, 20 object properties, 12 data properties and thousands of individuals divided into
80 test cases. In fact, in IIMB 2010, we have defined 80 test cases, divided into 4 sets of
20 test cases each. The first three sets are different implementations of data value, data
structure and data semantic transformations, respectively, while the fourth set is ob-
tained by combining together the three kinds of transformations. IIMB 2010 is created
by extracting data from Freebase, an open knowledge base that contains information
about 11 million real objects including movies, books, TV shows, celebrities, locations,
companies and more. Data extraction has been performed using the query language
JSON together with the Freebase JAVA API11. The benchmark has been generated in a
small version consisting in 363 individuals and in a large version containing 1416 indi-
viduals. In Figures 9 and 10, we report the results over the large version that are quite
similar to the small one.
111
!"##$ !"#%$ !"#&$ !"##$
!"&%$!"#’$
’"!!$ ’"!!$ ’"!!$
!"#’$
!"%($!"&!$
!"#)$
!"%#$!"&*$
!"&+$
!"%($!"%%$!"%#$
’"!!$
!"#!$!"#)$
!"##$ !"##$
!"&($
!"#&$!"#+$
!"*($!"),$ !")&$
!"%&$
!"*)$
!",($
!"**$ !"*%$ !"*&$
!"!!$
!"’!$
!"+!$
!"(!$
!")!$
!"*!$
!",!$
!"%!$
!"&!$
!"#!$
’"!!$
-./012134$ ./0566$ 789 /52:./$ -./012134$ ./0566$ 789 /52:./$ -./012134$ ./0566$ 789 /52:./$
;<= >?$ @>AB$ C1= >= $
?56:/$D.542E3.9 5F34$ <G.:0G:./$D.542E3.9 5F34$ H3I1056$D.542E3.9 5F34$ @39 -./J/421K/$$
Fig. 9. Results of the IIMB subtrack.
The participation in IIMB was limited to ASMOV, CODI and RiMOM systems. All
the systems obtained very good results when dealing with data value transformations
and logical transformations, both in terms of precision and in terms of recall. Instead,
in case of structural transformations (e.g., property value deletion of addition, property
hierarchy modification) and of the combination of different kinds of transformations
we have worse results, especially concerning recall. Looking at the results, it seems
that the combination of different kinds of heterogeneity in data descriptions is still an
open problem for instance matching systems. The three matching systems appear to be
comparable in terms of quality of results.
PR. The Person-Restaurants benchmark is composed of three data subsets. Two
datasets (Person 1 and Person 2) contain personal data. The Person 1 dataset is cre-
ated with the help of the Febrl project example datasets12. It contains original records
of people and modified duplicate records of the same entries. The duplicate record set
contains one duplicate per original record, with a maximum of one modification per
duplicate record and a maximum of one modification per attribute. Person 2 is created
as Person 1, but with a maximum of 3 modifications per attribute, and a maximum
of 10 modifications per record. The third dataset (Restaurant) is created with the help
of 864 restaurant records from two different data sources (Fodor and Zagat restaurant
guides)13. Restaurants are described by name, street, city, phone and restaurant cate-
gory. In all the datasets the number of records is quite limited (about 500/600 entries).
Among these, 112 record pairs refer to the same entity, but usually show differences.
Results of the evaluation are shown in Figure 11.
The PR task of the instance matching track was quite successful in terms of partici-
pation, in that all the five systems sent their results for this task14. This is because the PR
datasets contain a small number of instances to be matched, resulting in a matching task
11 http://code.google.com/p/freebase-java/12 Downloaded from http://sourceforge.net/projects/febrl/13 They can be downloaded from http://userweb.cs.utexas.edu/users/ml/riddle/data.html
14 ASMOV sent a second set of results referred as ASMOV D. They are the same as ASMOV
but alignments are generated using the descriptions available in the TBOX
112
Fig. 10. Precision/recall of tools participating in the IIMB subtrack.
that is affordable in terms of time required for comparisons. The results are good for
all systems with the best performances obtained by RiMOM followed by ObjectCoref
and LN2R. ASMOV and CODI instead have lower F-measure values in the case of the
Person 2 dataset. This is mainly due to low performances in terms of recall. These low
recall values depend on the fact that in Person 2 more than one matching counterpart
was expected for each person record in the reference dataset.
8 Lesson learned and suggestions
We have seriously implemented the promises of last year with the provision of the first
automated service for evaluating ontology matching, the SEALS evaluation service,
which has been used for three different data sets. We will continue on this path. We also
took into account two other lessons: having rules for submitting data sets and rules for
declaring them unfruitful that are published on the OAEI web site. There still remain
one lesson not really taken into account that we identify with an asterisk (*) and that
we will tackle next year.
The main lessons from this year are:
A) We were not sure that switching to an automated evaluation would preserve the
success of OAEI, given that the effort of implementing a web service interface was
required from participants. However, we still have as many participants as last year,
so this is a good sign.
113
Fig. 11. Results of tools participating in the PR subtrack in terms of F–measure.
B) Although some tools were registered in the SEALS portal, these have not used the
SEALS evaluation service either for testing their tools or for registering their final
results. We contacted these developers, who answered that they did not have enough
time for preparing their tools. So, the effort required for implementing the web
service interface and fixing networks issues, has indeed been an obstacle for some
participants. However, this effort is low with respect to that required for developing
a serious matcher.
C) The SEALS service eases the evaluation execution on a short period because par-
ticipants can improve their systems and obtain results in real time. This is to some
degree also possible for a blind evaluation. This is very valuable.
D) The trend that there are more matching systems able to enter such an evaluation
seems to slow down. There have been not many new systems this year but on spe-
cialized topics. There can be two explanations: the field is shrinking or the entry
ticket is too high. To address the first issue we have identified in [17] the challenges
in the field to direct research into the critical path.
E) We still can confirm that systems that enter the campaign for several times tend to
improve over years. We can also remark that they continue to improve (on data sets
in which there is still a progress margin).
F*) The benchmark test case is not discriminant enough between systems. Next year,
we plan to introduce controlled automatic test generation in the SEALS evaluation
service and think that this will improve the situation.
G) Not all systems followed the general rule to use the same set of parameters in all
tracks. In addition, there are systems participating only in one track for which they
are specialized. A fair comparison of general-purpose systems, specialized systems
and optimally configured systems might require to rethink the application of this
rule.
114
9 Future plans
There are several plans for improving OAEI. The first ones are related to the develop-
ment of the SEALS service. In the current setting, runtime and memory consumption
cannot be correctly measured because a controlled execution environment is missing.
Further versions of the SEALS evaluation service will include the deployment of tools
in such a controlled environment. As initially planned for last year, we plan to sup-
plement the benchmark test with an automatically generated benchmark that would be
more challenging for participants. We also plan to generalize the use of the platform to
other data sets. Finally, we would like to have again a data set which requires alignments
containing other relations than equivalence.
10 Conclusions
Confirming the trend of previous years, the number of systems, and tracks they enter in,
seems to stabilize. As noticed the previous years, systems which do not enter for the first
time are those which perform better. This shows that, as expected, the field of ontology
matching is getting stronger (and we hope that evaluation has been contributing to this
progress).
The trend of number of tracks entered by participants went down again: 2.6 against
3.25 in 2009, 3.84 in 2008 and 2.94 in 2007. This figure of around 3 out of 8 may be
the result of the specialization of systems. While, it is not the result of the short time
allowed to the campaign, since the SEALS evaluation service had more runs than what
the participants registered.
All participants have provided a description of their systems and their experience in
the evaluation. These OAEI papers, like the present one, have not been peer reviewed.
However, they are full contributions to this evaluation exercise and reflect the hard work
and clever insight people put in the development of participating systems. Reading the
papers of the participants should help people involved in ontology matching to find what
makes these algorithms work and what could be improved. Sometimes participants offer
alternate evaluation results.
The Ontology Alignment Evaluation Initiative will continue these tests by improv-
ing both test cases and testing methodology for being more accurate. Further informa-
tion can be found at:
http://oaei.ontologymatching.org.
AcknowledgmentsWe warmly thank each participant of this campaign. We know that they have worked
hard for having their results ready and they provided insightful papers presenting their
experience. The best way to learn about the results remains to read the following papers.
We also warmly thank Laura Hollinck, Veronique Malaise and Willem van Hage
for preparing the vlcr task which has been cancelled.
We are grateful to Martin Ringwald and Terry Hayamizu for providing the reference
alignment for the anatomy ontologies and thank Elena Beisswanger for her thorough
support on improving the quality of the data set.
115
We are grateful to Dominique Ritze for participating in the extension of the refer-
ence alignments for the conference track.
We thank Jan Noessner for providing data in the process of constructing the IIMB
dataset and we thank Heiko Stoermer and Nachiket Vaidya for providing the PR dataset
for Instance Matching.
We also thank the other members of the Ontology Alignment Evaluation Initiative
Steering committee: Yannis Kalfoglou (Ricoh laboratories, UK), Miklos Nagy (The
Open University (UK), Natasha Noy (Stanford University, USA), Yuzhong Qu (South-
east University, CN), York Sure (Leibniz Gemeinschaft, DE), Jie Tang (Tsinghua Uni-
versity, CN), George Vouros (University of the Aegean, GR).
Jerome Euzenat, Christian Meilicke, Heiner Stuckenschmidt and Cassia Trojahn
dos Santos have been partially supported by the SEALS (IST-2009-238975) European
project.
Ondrej Svab-Zamazal and Vojtech Svatek were supported by the CSF grant
P202/10/1825 (PatOMat project).
References
1. Benhamin Ashpole, Marc Ehrig, Jerome Euzenat, and Heiner Stuckenschmidt, editors. In-tegrating Ontologies’05, Proc. of the K-Cap Workshop on Integrating Ontologies, Banff
(Canada), 2005.
2. Oliver Bodenreide, Terry Hayamizu, Martin Ringwald, Sherri De Coronado, and Songmao
Zhang. Of mice and men: Aligning mouse and human anatomies. In Proc. of the AmericanMedical Informatics Association (AIMA) Annual Symposium, pages 61–65, 2005.
3. Caterina Caracciolo, Jerome Euzenat, Laura Hollink, Ryutaro Ichise, Antoine Isaac,
Veronique Malaise, Christian Meilicke, Juan Pane, Pavel Shvaiko, Heiner Stuckenschmidt,
Ondrej Svab-Zamazal, and Vojtech Svatek. Results of the ontology alignment evaluation ini-
tiative 2008. In Proc. of the 3rd International Workshop on Ontology Matching (OM-2008),collocated with ISWC-2008, pages 73–120, Karlsruhe (Germany), 2008.
4. Marc Ehrig and Jerome Euzenat. Relaxed precision and recall for ontology matching. In
Proc. of the K-Cap Workshop on Integrating Ontologies, pages 25–32, Banff (Canada), 2005.
5. Jerome Euzenat. An API for ontology alignment. In Proc. of the 3rd International SemanticWeb Conference (ISWC-2004), pages 698–712, Hiroshima (Japan), 2004.
6. Jerome Euzenat, Alfio Ferrara, Laura Hollink, Antoine Isaac, Cliff Joslyn, Veronique
Malaise, Christian Meilicke, Andriy Nikolov, Juan Pane, Marta Sabou, Francois Scharffe,
Pavel Shvaiko, Vassilis Spiliopoulos, Heiner Stuckenschmidt, Ondrej Svab-Zamazal, Vo-
jtech Svatek, Cassia Trojahn dos Santos, George Vouros, and Shenghui Wang. Results of
the ontology alignment evaluation initiative 2009. In Proc. of the 4th Workshop on OntologyMatching (OM-2009), collocated with ISWC-2000, pages 73–126, Chantilly (USA), 2009.
7. Jerome Euzenat, Antoine Isaac, Christian Meilicke, Pavel Shvaiko, Heiner Stuckenschmidt,
Ondrej Svab, Vojtech Svatek, Willem Robert van Hage, and Mikalai Yatskevich. Results of
the ontology alignment evaluation initiative 2007. In Proc. of the 2nd International Work-shop on Ontology Matching (OM-2008), collocated with ISWC-2007, pages 96–132, Busan
(Korea), 2007.
8. Jerome Euzenat, Malgorzata Mochol, Pavel Shvaiko, Heiner Stuckenschmidt, Ondrej Svab,
Vojtech Svatek, Willem Robert van Hage, and Mikalai Yatskevich. Results of the ontology
alignment evaluation initiative 2006. In Proc. of the 1st International Workshop on Ontology
116
Matching (OM-2006), collocated with ISWC-2006, pages 73–95, Athens, Georgia (USA),
2006.
9. Jerome Euzenat and Pavel Shvaiko. Ontology Matching. Springer, Heidelberg (DE), 2007.
10. Fausto Giunchiglia, Mikalai Yatskevich, Paolo Avesani, and Pavel Shvaiko. A large scale
dataset for the evaluation of ontology matching systems. The Knowledge Engineering ReviewJournal, 24(2):137–157, 2009.
11. Oktie Hassanzadeh, Reynold Xin, Renee J. Miller, Anastasios Kementsietsidis, Lipyeow
Lim, and Min Wang. Linkage query writer. PVLDB, 2(2):1590–1593, 2009.
12. Yves R. Jean-Mary, E. Patrick Shironoshitaa, and Mansur R. Kabuka. Ontology matching
with semantic verification. Journal of Web Semantics, 7(3):235–251, 2009.
13. Anja Jentzsch, Jun Zhao, Oktie Hassanzadeh, Kei-Hoi Cheung, Matthias Samwald, and
Bo Andersson. Linking open drug data. In Proc. of Linking Open Data Triplification Chal-lenge at the I-Semantics, 2009.
14. Christian Meilicke and Heiner Stuckenschmidt. Incoherence as a basis for measuring the
quality of ontology mappings. In Proc. of the 3rd Workshop on Ontology Matching (OM-2008), collocated ISWC-2008, pages 1–12, Karlsruhe (Germany), 2008.
15. Vojtech Svatek Ondrej Svab-Zamazal O. Empirical knowledge discovery over ontology
matching results. In Proc. of the 1st International Workshop on Inductive Reasoning andMachine Learning on the Semantic Web (IRMLeS-2009), collocated with ESWC-2009, pages
1–15, Heraklion (Greece), 2009.
16. Francois Scharffe. Correspondence Patterns Representation. PhD thesis, University of Inns-
bruck, 2009.
17. Pavel Shvaiko and Jerome Euzenat. Ten challenges for ontology matching. In Proc. ofthe 7th International Conference on Ontologies, DataBases, and Applications of Semantics(ODBASE-2008), pages 1164–1182, Monterrey (Mexico), 2008.
18. York Sure, Oscar Corcho, Jerome Euzenat, and Todd Hughes, editors. Proc. of the Workshopon Evaluation of Ontology-based Tools (EON-2004), collocated with ISWC-2004, Hiroshima
(Japan), 2004.
19. Cassia Trojahn dos Santos, Christian Meilicke, Jerome Euzenat, and Heiner Stuckenschmidt.
Automating OAEI campaigns (first report). In Proc. of the 1st International Workshop onEvaluation of Semantic Technologies (iWEST-2010), collocated with ISWC-2010, Shanghai
(China), 2010.
20. Willem Robert van Hage, Antoine Isaac, and Zharko Aleksovski. Sample evaluation of
ontology-matching systems. In Proc. of the 5th International Workshop on Evaluation ofOntologies and Ontology-based Tools (EON-2007), collocated with ISWC-2007, pages 41–
50, Busan (Korea), 2007.
21. Julius Volz, Christian Bizer, Martin Gaedke, and Georgi Kobilarov. Discovering and main-
taining links on the web of data. In Proc. of the 8th International Semantic Web Conference(ISWC-2009), pages 650–665, Chantilly (USA), 2009.
22. Ondrej Svab, Vojtech Svatek, and Heiner Stuckenschmidt. A study in empirical and ‘ca-
suistic’ analysis of ontology mapping results. In Proc. of the 4th European Semantic WebConference (ESWC-2007), pages 655–669, Innsbruck (Austria), 2007.
Grenoble, Milano, Mannheim, Milton-Keynes, Trento, Prague, December 2010
117
Using AgreementMaker to Align Ontologies for OAEI2010�
Isabel F. Cruz, Cosmin Stroe, Michele Caci, Federico Caimi, Matteo Palmonari��,
Flavio Palandri Antonelli, Ulas C. Keles
ADVIS Lab, Department of Computer Science, University of Illinois at Chicago
{ifc|cstroe1|mcaci|fcaimi|pmatteo|flav|ukeles}@cs.uic.edu
Abstract. The AgreementMaker system is unique in that it features a powerful
user interface, a flexible and extensible architecture, an integrated evaluation en-
gine that relies on inherent quality measures, and semi-automatic and automatic
methods. This paper describes the participation of AgreementMaker in the 2010
OAEI competition in three tracks: benchmarks, anatomy, and conference. After
its successful participation in 2009, where it ranked first in the conference track,
second in the anatomy track, and obtained good results in the benchmarks track,
the goal in this year’s participation is to increase the values of precision, recall,
and F-measure for each of those tracks.
1 Presentation of the system
We have been developing the AgreementMaker system since 2001, with a focus on
real-world applications [5, 8] and in particular on geospatial applications [4, 6, 7, 9–13].
However, the current version of AgreementMaker, whose development started two years
ago, represents a whole new effort.
1.1 State, purpose, general statement
The new AgreementMaker system [1–3] supports: (1) user requirements, as expressed
by domain experts; (2) a wide range of input (ontology) and output (agreement file)
formats; (3) a large choice of matching methods depending, on the different granularity
of the set of components being matched (local vs. global), on different features consid-
ered in the comparison (conceptual vs. structural), on the amount of intervention that
they require from users (manual vs. automatic), on usage (standalone vs. composed),
and on the types of components to consider (schema only or schema and instances); (4)
improved performance, that is, accuracy (precision, recall, F-measure) and efficiency
(execution time) for the automatic methods; (5) an extensible architecture to incorpo-
rate new methods easily and to tune their performance; (6) the capability to evaluate,
compare, and combine different strategies and matching results; (7) a comprehensive
user interface that supports advanced visualization techniques and a control panel that
� Research supported by NSF Awards IIS-0513553 and IIS-0812258.�� Additional affiliation: University of Milan-Bicocca, Italy.
118
drives all the matching methods and evaluation strategies; (8) a feedback loop that ac-
cepts suggestions and corrections by users and extrapolates new mappings.
In 2009 AgreementMaker was very successful in the OAEI competition. In par-
ticular, AgreementMaker ranked (a close) second among ten systems in the anatomy
track. AgreementMaker also participated successfully in two other tracks: benchmarks
and conference. In the former track, AgreementMaker was ranked first in terms of pre-
cision and seventh in terms of recall among thirteen systems and in the latter track
AgreementMaker was ranked first with the highest F-measure (57% at a threshold of
75%) among seven competing systems.
1.2 Specific techniques used
AgreementMaker comprises several matching algorithms or matchers that can be used
for matching (or aligning) the source and target ontologies. The matchers are not re-
stricted to any particular domain. The architecture of AgreementMaker relies on a stack
of matchers that belong to three different layers (see Figure 1). Specific configurations
of the stack have been used for the benchmarks, anatomy, and conference tracks, as
discussed in what follows. However, we describe first the different components in the
stack: the matchers, the combination and evaluation modules, and the final alignment
module.
Fig. 1. AgreementMaker OAEI 2010 matcher stack.
119
Matchers can be concept-based (if they consider only one concept) or structural
(if they consider a subgraph of the ontology). The concept-based matchers support the
comparison of strings. They include: the Base Similarity Matcher (BSM) [7], the Para-
metric String-based Matcher (PSM) [2] and the Vector-based Multi-Word
Matcher (VMM) [2]. BSM is a basic string matcher that computes the similarity be-
tween concepts by comparing all the strings associated with them. PSM is a more
in-depth string matcher, which for the competition is set to use a substring measure
and an edit distance measure. VMM compiles a virtual document for every concept
of an ontology, transforms the resulting strings into TF-IDF vectors and then computes
their similarity using the cosine similarity measure. These matchers have been extended
in the AgreementMaker configuration used this year by plugging in a set of lexicons,
which are used to expand the set of strings with synonyms. The extended matchers
are therefore called BSMlex, PSMlex, and VMMlex. The Advanced Similarity Matcher
(ASM) is a string-based matcher that computes mappings between source and target
concepts (including their properties) by comparing their local names, and providing
better similarity evaluation in particular when compound terms are used. ASM outper-
forms generic string-based similarity matchers because it is based on a deeper linguistic
analysis.
Structural matchers include the Descendants’ Similarity Inheritance (DSI)
matcher [7]. This matcher is based on the idea that if two nodes are similar, then their
descendants should be similar. The Group Finder Matcher (GFM) is another structural
matcher that filters out the mappings provided by another matcher (the input matcher).
It identifies groups of concepts and properties in the ontologies and assumes that two
concepts (or properties) that belong to two groups that were not mapped by the in-
put matcher will likely have different meanings and should not be mapped. The Itera-
tive Instance Structural Matcher (IISM) takes into account instances. Classes that have
mapped individuals can then be aligned. In addition, values of the properties are also
considered. The structural part of IISM is quite complex and takes into account super-
classes, subclasses, properties, subproperties, cardinalities, and the range and domain
of properties.
The combination and evaluation modules are used together, as follows. The Linear
Weighted Combination (LWC) [2] combines its inputs (e.g., from several string match-
ers), using a local confidence quality measure provided by the evaluation module, in
order to automatically assign weights to each result computed by the input matchers.
After this step, we have a single combined set of alignments that includes the best align-
ments from each of the input matchers. The final alignment module is given as input a
mapping cardinality (e.g., 1:1) and a threshold and outputs the best set of alignments
given those two inputs [2].
Benchmarks For the benchmarks track we used the following configuration:
IISM( LWC(ASM, PSMlex,VMMlex,BSMlex) )
LWC is adopted to combine the results of four string-based matchers, namely ASM,
PSMlex, PSMlex, and BSMlex; the last three make use of two lexicons, namely Word-
Net and a dictionary built from the ontologies; the similarity values computed at this
step are then given as input to the IISM structural matcher.
120
Anatomy For the anatomy track we used the following configuration:
LWC(PSMlex,VMMlex,BSMlex)
LWC is adopted to combine the results of four string-based matchers, namely PSMlex,
VMMlex, and BSMlex; the last three make use of two lexicons, namely WordNet and a
dictionary built from the ontologies.
Conference For the conference track we used the following configuration:
GFM( LWC(ASM, PSM) )
LWC is adopted to combine the results of two string-based matchers, namely ASM
and PSM; the similarity values computed at this step are then given as input to the GFM
structural matcher.
1.3 Link to the system and parameters file
The AgreementMaker system is available at http://agreementmaker.org/.
1.4 Link to the set of provided alignment (in align format)
The alignment results obtained by AgreementMaker in the OAEI 2010 are available at
http://agreementmaker.org/oaei.
2 Results
In this section, we present the results obtained by AgreementMaker in the OAEI 2010
competition. It participated in three tracks: benchmarks, anatomy, and conference. Tests
were carried out on a PC running Ubuntu Linux 10.04 with AMD AthlonTM II X4 635
processor running at 2.9 Ghz and 8 GB RAM.
2.1 Benchmarks
In this track, a source ontology is compared to 111 ontologies that describe the same
domain. These ontologies can be divided into 3 categories: concept tests cases (1xx
cases), systematic tests cases (2xx cases), and real ontology test cases (3xx cases).
AgreementMaker employs the algorithm which we described in section 1.2 for aligning
two ontologies.
The 2xx benchmarks test cases are subdivided into 3 groups: 1) 201 to 210, 2)
221 to 247 and 3) 248 to 266. The lexical information in the ontologies in group 1
have been altered to change their labels or identifiers. This alteration includes replacing
the labels or identifiers with other names that follow a particular naming convention,
a random name, a misspelled name or a foreign word. However, the structure of the
ontologies is not modified. The test cases in the second group have ontologies that have
flattened hierarchies, expanded hierarchies or no hierarchies at all. The test cases in
the third group are the most challenging ones to align. This is because the labels have
121
been scrambled such that they comprise a permutation of letters of a particular length.
In addition, the structure of the ontology has been flattened, expanded such that it has
more depth or removed completely.
The results obtained by AgreementMaker in the benchmarks track are summarized
in Table 1.
101-104 201-210 221-247 248-266 301-304 H-mean
precision 0.98 0.97 0.95 0.96 0.88 0.95
recall 1.00 0.90 0.99 0.74 0.53 0.79
F-measure 0.99 0.94 0.97 0.82 0.61 0.84
Table 1. Results achieved by AgreementMaker in the benchmarks track of the OAEI 2010
competition.
2.2 Anatomy
This track consists of two real world ontologies to be matched. The source ontology de-
scribes the Adult Mouse Anatomy (with 2744 classes) while the target ontology is the
NCI Thesaurus describing the Human Anatomy (with 3304 classes). Matching these on-
tologies is also challenging in terms of efficiency because these ontologies are relatively
large. The anatomy track consists of four subtracks: subtrack 1, which emphasizes F-
measure, subtrack 2, which emphasizes precision, subtrack 3, which emphasizes recall,
and subtrack 4, which tests the capability of extending a partial reference alignment.
The results obtained by AgreementMaker in the anatomy track are summarized in
Table 2. We show the precision, recall, and F-measure for subtrack 1, 2 and 3; subtracks
1 and 3 are also evaluated by considering the recall+ measure, which measures how
many non trivial correct correspondences, not detectable by string equivalence, can be
found in an alignment. Evaluation of subtrack 4 is carried out by analyzing the changes
in precision, recall and f-measure when subtrack 1 is compared with subtrack 4.
Anatomy Track Subtrack 1 Subtrack 2 Subtrack 3 Subtrack 4
precision 0.90 0.96 0.77 +0.03
recall 0.85 0.75 0.87 -0.03
F-measure 0.87 0.84 0.82 +0.00
recall+ 0.63 - 0.70 -
Table 2. Results achieved by AgreementMaker in the anatomy track of the OAEI 2010 compe-
tition.
2.3 Conference
The conference track consists of 15 ontologies from the conference organization do-
main and each ontology must be matched against every other ontology. Since the
AgreementMaker OAEI 2010 matcher stack considers only two ontologies at a time,
we compute 120 alignment files, in total containing 2070 individual alignments. The
results obtained are summarized in Table 3. Here we show how precision, recall, and
F-measure vary depending on the threshold used for the selection of the mappings.
122
Threshold 0.60-0.64 0.65-0.69 0.70-0.74 0.75-0.79 0.80-0.84 0.85-0.89 0.90-0.94 0.95-1.00
precision 0.49 0.53 0.61 0.70 0.72 0.75 0.81 0.81
recall 0.64 0.60 0.54 0.52 0.47 0.47 0.46 0.46
F-measure 0.56 0.56 0.58 0.60 0.58 0.58 0.58 0.59
Table 3. Results achieved in the conference track of the OAEI 2010 competition.
2.4 Comments on the results
Benchmarks In the OAEI 2009 competition, AgreementMaker was first in terms of
the precision of discovered mappings. However, in terms of recall, AgreementMakerwas outperformed by six other systems (thirteen systems competed). The new matchers
used in the OAEI 2010 competition address specifically the issue of the alignment of
concepts that are not lexically similar. The results of this effort increased the recall by
18% at a cost of 3% in precision in comparison with last year’s results. An important
contribution to this result comes from the IISM matcher, which exploits instances and
structural properties of the ontologies and makes the alignment process less sensitive to
lexical differences. A detailed comparison between the results achieved in the 2009 and
2010 competitions in terms of the obtained change in precision, recall, and F-measure
for each group of test cases, and the overall H-mean is shown in Table 4.
101-104 201-210 221-247 248-266 301-304 H-mean
precision 09 0.86 0.73 0.76 0.47 0.83 0.70
precision 10 0.98 0.97 0.95 0.96 0.88 0.95
recall 09 0.86 0.73 0.76 0.47 0.86 0.70
recall 10 1.00 0.90 0.99 0.74 0.53 0.79
F-measure 09 0.92 0.71 0.86 0.45 0.83 0.70
F-measure 10 0.99 0.94 0.97 0.82 0.61 0.84
Table 4. Comparison of the results achieved by AgreementMaker in the 2009 and 2010 OAEI
benchmarks track.
Anatomy In comparison with the results achieved by AgreementMaker in the OAEI
2009 competition, the experimental results obtained this year show that the system sig-
nificantly improved with respect to precision, recall, F-measure, and recall+. A major
contribution to these results comes from the exploitation of lexical resources to improve
string-based and vector-based matchers. A comparison between the results achieved in
the two competitions in terms of precision, recall, F-measure and recall+ for subtracks
1, 2 and 3 is shown in Table 5 (except for recall+ that is not evaluated on subtrack 2).
Remarkably, our algorithms for retrieving non trivial mappings significantly improved,
as shown by the gain of 0.15 in recall+. Instead, we do not present the comparison with
the results obtained in 2009 on subtrack 4, because this year we did not exploit any
specific algorithm for propagating mappings available in the input alignment.
123
Anatomy Track Subtrack 1 Subtrack 2 Subtrack 3
precision 09 0.87 0.97 0.51
precision 10 0.90 0.96 0.77
recall 09 0.80 0.68 0.82
recall 10 0.85 0.75 0.87
F-measure 09 0.83 0.80 0.63
F-measure 10 0.87 0.84 0.82
recall+ 09 0.49 - 0.55
recall+ 10 0.63 - 0.70
Table 5. Comparison of the results achieved by AgreementMaker in the 2009 and 2010 OAEI
anatomy track.
Conference In comparison with the results achieved in OAEI 2009,
AgreementMaker significantly improved on precision, recall, and F-measure for thresh-
olds above 0.75 as shown in the graph represented of Figure 2.4, providing more stable
results. Remarkably, the new matchers used for the conference track, namely ASM and
GFM, can be used on real-world ontologies, since they are based on generic lexical
and structural features. Moreover, ASM can be easily adapted to different string-based
similarity metrics, and can be extended by adopting a lexicon.
Fig. 2. F-measure comparison for the 2009 and 2010 OAEI conference track results.
3 Conclusions
In this paper we presented the results of the AgreementMaker system for aligning on-
tologies in the OAEI 2010 competition in the three tracks in which it participated:
benchmarks, anatomy, and conference. It was our goal to improve on the results ob-
tained by AgreementMaker in 2009. To meet this goal, we developed several new match-
124
ing methods, which could be readily integrated into the AgreementMaker system be-
cause of its modularity and extensibility. Our results (which we compare with last year’s
results) amply demonstrate that we have met our goal.
Acknowledgements
Thanks to Catia Pesquita and Francisco M. Couto from Facultade de Ciencias da Uni-
versitade de Lisboa, Portugal, for the feedback on the AgreementMaker system, and for
useful insights on biomedical ontologies, which significantly increased our understand-
ing of the problem and our results in the anatomy track.
References
1. Isabel F. Cruz, Flavio Palandri Antonelli, and Cosmin Stroe. AgreementMaker: Efficient
Matching for Large Real-World Schemas and Ontologies. PVLDB, 2(2):1586–1589, 2009.
2. Isabel F. Cruz, Flavio Palandri Antonelli, and Cosmin Stroe. Efficient Selection of Mappings
and Automatic Quality-driven Combination of Matching Methods. In ISWC InternationalWorkshop on Ontology Matching. CEUR-WS, 2009.
3. Isabel F. Cruz, Flavio Palandri Antonelli, and Cosmin Stroe. Integrated Ontology Matching
and Evaluation. In International Semantic Web Conference (Posters & Demos), 2009.
4. Isabel F. Cruz and Afsheen Rajendran. Exploring a New Approach to the Alignment of
Ontologies. In ISWC Workshop on Semantic Web Technologies for Searching and RetrievingScientific Data, volume 83 of CEUR-WS, pages 7–12, 2003.
5. Isabel F. Cruz and Afsheen Rajendran. Semantic Data Integration in Hierarchical Domains.
IEEE Intelligent Systems, March-April:66–73, 2003.
6. Isabel F. Cruz, Afsheen Rajendran, William Sunna, and Nancy Wiegand. Handling Semantic
Heterogeneities Using Declarative Agreements. pages 168–174, 2002.
7. Isabel F. Cruz and William Sunna. Structural Alignment Methods with Applications to
Geospatial Ontologies. Transactions in GIS, special issue on Semantic Similarity Measure-ment and Geospatial Applications, 12(6):683–711, 2008.
8. Isabel F. Cruz, William Sunna, and Anjli Chaudhry. Ontology Alignment for Real-World
Applications. In National Conference on Digital Government Research (dg.o), pages 393–
394, 2004.
9. Isabel F. Cruz, William Sunna, and Anjli Chaudhry. Semi-Automatic Ontology Alignment
for Geospatial Data Integration. volume 3234 of Lecture Notes in Computer Science, pages
51–66. Springer, 2004.
10. Isabel F. Cruz, William Sunna, Nalin Makar, and Sujan Bathala. A Visual Tool for On-
tology Alignment to Enable Geospatial Interoperability. Journal of Visual Languages andComputing, 18(3):230–254, 2007.
11. Isabel F. Cruz, William G. Sunna, and Kalyan Ayloo. Concept Level Matching of Geospatial
Ontologies. 2005.
12. Isabel F. Cruz and Huiyong Xiao. Data Integration for Querying Geospatial Sources. In John
Sample, Kevin Shaw, Shengru Tu, and Mahdi Abdelguerfi, editors, Geospatial Services andApplications for the Internet, pages 113–137. Springer, 2008.
13. William Sunna and Isabel F. Cruz. Structure-Based Methods to Enhance Geospatial Ontol-
ogy Alignment. In International Conference on GeoSpatial Semantics (GeoS), volume 4853
of Lecture Notes in Computer Science, pages 82–97. Springer, 2007.
125
ASMOV: Results for OAEI 2010
Yves R. Jean-Mary1, E. Patrick Shironoshita1, Mansur R. Kabuka 1,2
1 INFOTECH Soft, 1201 Brickell Ave., Suite 220, Miami, Florida, USA 331312 University of Miami, Coral Gables, Florida, USA 33124
{reggie, kabuka}@infotechsoft.com
Abstract. The Automated Semantic Mapping of Ontologies with Validation(ASMOV) algorithm for ontology alignment has consistently been one of the top performing algorithms in the Ontology Alignment Evaluation Initiative (OAEI) contests. In this paper, we present a brief overview of the algorithm and its improvements, followed by an analysis of its results on the 2010 OAEI tests.
1 Presentation of the System
In recent years, ontology alignment has become popular in solving interoperability issues across heterogonous systems in the semantic web. There exist many techniques to address this problem [1], differentiated by the way in which different ontology features are exploited. ASMOV, an algorithm that automates the ontology alignment process, uses a weighted average of measurements of similarity along four different features of ontologies, and obtains a pre-alignment based on these measurements. It then uses a unique process of semantic verification to ensure that the alignment does not contain semantic inconsistencies. In this manner, ASMOV was shown to produce the most coherent alignments of all systems tested in OAEI 2009 [3]. A more complete description of ASMOV is presented in [4].
1.1 State, Purpose, General Statement
ASMOV is an automatic ontology matching tool which has been designed in order to facilitate the integration of heterogeneous data sources modeled as ontologies. The current ASMOV implementation produces mappings between concepts, properties, and individuals, including mappings between object and datatype properties.
1.2 Specific Techniques Used
The ASMOV algorithm iteratively calculates the similarity between entities for a pair of ontologies by analyzing four features: lexical elements (id, label, and comments), relational structure (ancestor-descendant hierarchy), internal structure (property restrictions for concepts; types, domains, and ranges for properties; data values for individuals), and extension (instances of classes and property values). The measures
126
obtained by comparing these four features are combined into a single value using a weighted sum in a similar manner to [2]. These weights have been optimized based on the OAEI 2008 benchmark test results.
Fig. 1. The ASMOV Mapping Process
Fig. 1 illustrates the fully automated ASMOV mapping process, which has been implemented in Java. In the pre-processing phase, the ontologies are loaded into memory using the Jena ARP parser [5] and ASMOV’s ontology modeling component. A thesaurus is optionally used to calculate the lexical similarities between each pair of concepts, properties and individuals. ASMOV can be configured to use either the UMLS Metathesaurus [6] or WordNet [7] in order to derive the similarity measures. If a thesaurus is not used, a text matching algorithm is used to compute the lexical distance. Following this, the similarities between pairs of entities along the relational structure, internal structure, and extensional dimensions are calculated, and overall similarity measures (or confidence values) are calculated for each pair. From these similarity measures, a pre-alignment is obtained by selecting the entity from one ontology with the highest similarity for a corresponding entity in the other ontology.A threshold of 0.1% is used to ignore spurious non-zero similarity measures.
This pre-alignment then goes through semantic verification, which detects semantically inconsistent mappings and their causes. These inconsistent mappings are removed from the pre-alignment and logged so that the algorithm does not attempt to map the same entities in a subsequent iteration; mappings are removed from the log of inconsistencies when the underlying cause disappears. Five specific types of inconsistencies are detected by ASMOV:
Multiple entity correspondences, where the same entity on one ontology is mapped with multiple entities in the other ontology; unless these multiple entities are asserted to be equivalent, this type of mapping is unverified.Crisscross correspondences, where if a class c1 in one ontology is mapped to some other class c1‘ in the second ontology, a child of c1 cannot be mapped to a parent of c1‘.Disjointness-subsumption contradiction, where if two classes c1 and c2 are disjoint in one ontology, they cannot be mapped to two other classes c1‘ and c2‘ in the second ontology where one is subsumed by the other. This also
127
applies to the special cases where c1‘ and c2‘ are asserted equivalent, or where they are identical.Subsumption incompleteness, if two classes c1 and c2 are mapped to two other classes c1‘ and c2‘ respectively in the second ontology, and if c2 is subsumed by c1, then c2‘ must be subsumed by c1‘, otherwise the correspondences are unverified. Similar incompleteness can be verified for the special case of equivalence.Domain and range incompleteness: if a class c1 in one ontology is mapped to some class c1‘ in the second ontology, and a property p1 in the first ontology is mapped to some property p1‘ in the second ontology, and if c1 belongs to the domain (or range) of p1 , then c1‘ must belong to the domain (or, equivalently, range) of p1‘,
Since OAEI 2009, ASMOV has been improved in three important respects, primarily related to the new instance matching tests. The algorithm in general has been enhanced to allow it to process certain property constructs introduced in OWL 2, especially irereflexive and asymmetric properties. A procedure for disk-based storage of intermediate results has been implemented, allowing the algorithm to handle larger ontologies, although the ontology itself still needs to reside in memory. Further, we have improved the ability of ASMOV to use reasoning enabled by OWL in order to find semantically relevant matches. In particular, we have improved the verification of disjointness between domains and ranges of properties, and we also have included verification of functional properties.
1.3 Adaptations Made for the Evaluation
No special adaptations have been made to the ASMOV system in order to run the 2010 OAEI tests. The stop criterion for ASMOV was established as a multiple-alignment change threshold. For situations where both ontologies have more than 500 concepts, this threshold was set at 1% over three consecutive alignments; otherwise, it was set at 0% over two consecutive alignments. Although the rules of the contests stated that all alignments should be run from the same set of parameters, it was necessary to change two parameters for the anatomy tests. These parameters relate tothe thesaurus being used (UMLS instead of WordNet) and to the flag indicating whether or not to use ids of entities in the lexical similarity calculations.
1.4 Link to the ASMOV System
The ASMOV system (including the parameters file) can be downloaded from http://www.infotechsoft.com/products/asmov.aspx.
1.5 Link to the Set of Alignments Produced by ASMOV
The results of the 2010 OAEI campaign for the ASMOV system can be found at http://www.infotechsoft.com/products/asmov.aspx.
128
2 Results
In this section, we present our comments on the results obtained from the participation of ASMOV in the five tracks of the 2010 Ontology Alignment Evaluation Initiative campaign. Tests were carried out on a PC running FreeBSD over VMware with two quad-core Intel Xeon processor (1.86 GHz), 8 GB of memory, and 2x4MB cache, with a Web service interface to run with the SEALS environment where required.
2.1 Benchmark
The OAEI 2010 benchmark tests have been divided by the organizing committee in eleven levels of difficulty; we have added one more level to include the set of 3xx tests, which have been included in the benchmark for compatibility with previous years. In Table 1, we present the results of these tests in comparison with those obtained in 2009 [8], where ASMOV was found to be one of the two best performing systems [3]. As can be seen, the precision, recall, and F1 measure for the entire suite of tests shows that ASMOV 2010 achieves 98% precision and 89% recall, and an F1 measure of 93%, which represents a 2% improvement over the 2009 version.
The accuracy of ASMOV in the benchmark tests is very high, especially for the lowest levels of difficulty. It is particularly noteworthy that improvements in both precision and recall were obtained especially at higher levels, with the largest improvement within level 10, the most difficult, and with significant improvements at levels 7 through 9 and at the 3xx tests. We believe that these improvements have come specifically through the enhancement of the procedures for utilizing domain and range information in the calculation of internal structure similarity, and through the correction of coding errors and deficiencies. In addition, some of this improvement can be attributable to improvements made in the gold standard.
Table 1. Benchmark test results for ASMOV version 2010 and version 2009Level ASMOV 2010 ASMOV 2009
Precision Recall F1 Precision Recall F10 1.00 1.00 1.00 1.00 1.00 1.001 0.99 1.00 0.99 1.00 1.00 1.002 1.00 0.99 0.99 1.00 0.99 0.993 0.99 0.98 0.98 0.99 0.98 0.984 1.00 0.98 0.99 0.99 0.98 0.985 0.99 0.94 0.96 0.97 0.93 0.956 0.98 0.90 0.94 0.95 0.89 0.927 0.98 0.87 0.92 0.93 0.83 0.888 0.98 0.77 0.86 0.90 0.71 0.799 0.97 0.64 0.77 0.83 0.48 0.61
10 0.90 0.29 0.44 0.40 0.04 0.073xx 0.88 0.84 0.86 0.81 0.82 0.81All 0.98 0.89 0.93 0.95 0.87 0.91
129
2.2 Anatomy
For the anatomy track, ASMOV uses the UMLS Metathesaurus [6] instead of WordNet in order to more accurately compute the lexical distance between medical concepts. In addition, the lexical similarity calculation between concept names (ids) is ignored as instructed by the track organizers. ASMOV produces an alignment for all four subtasks of this track; the SEALS platform provides accuracy measurements for the first three subtasks.1. Optimal solution: The optimal solution alignment is obtained by using the default
parameter settings of ASMOV. The accuracy figures obtained from SEALS indicate precision of 79.9% and recall of 77.2%, resulting in overall F1 of 78.5%;these figures are a distinct improvement over the results obtained in 2009.
2. Optimal precision: The alignment with optimal precision is obtained by changing the threshold for valid mappings from 0.1% to 30%. The result is that precision increases to 86.5%, while recall decreases to 75.7%. F1 measure is 80.8%, which is higher than our optimal solution, indicating that the use of a higher threshold for ASMOV should be studied more closely.
3. Optimal recall: To improve recall, this time ASMOV made use of the annotation property hasRelatedSynonym included in the ontologies, to signify synonyms. It should be emphasized that this property is not included in the optimal solution because annotation properties do not have established semantics in OWL, and therefore it would not be possible for a computer to automatically understand that this property actually lists synonyms The results from SEALS indicate that ASMOV found a total of 1521 alignments with precision of 71.7%and recall of 79.2%, resulting in F1 of 75.3%.
4. Extended solution: With a partial alignment given as input, the resulting alignment contained all mappings in the partial plus an additional 480 mappings.
2.3 Conference
This collection of tests dealing with conference organization contains 16 ontologies,of which at least one contains constructs specific to OWL 2. ASMOV is able to generate all 120 potential alignments from those ontologies.
Our analysis of the preliminary results obtained in running ASMOV against these ontologies showed a large number of erroneous matches due to incompleteness in our processing of disjointness between property domains and ranges. Specifically, our
Table 2. Results for Conference Test
F1 cmt confer. Confof edas ekaw iasted sigkddcmt 0.476 0.378 0.556 0.437 0.364 0.541confer. 0.718 0.453 0.451 0.286 0.500confof 0.549 0.681 0.378 0.357edas 0.529 0.386 0.510ekaw 0.348 0.368iasted 0.481
130
previous versions of ASMOV only verified whether disjointness axioms existed in the asserted domain and range classes. We have now expanded ASMOV to verify any inferred disjointness between domains or ranges based on asserted disjointness within the subsumption hierarchy. Table 2 shows the results that were obtained by running the test through the SEALS platform.
2.4 Directory
Both the “small task” and the “single task” modalities were completed using ASMOV. The “small task” modality comprised 4639 tasks. We compared the results obtained this year against those obtained in 2009, where ASMOV was the best performing system [3]. We found a large degree of agreement, measured as 97% F1. We believe that the difference should result in improvement in the performance of ASMOV over 2009. The “single task” modality consisted of the alignment of a source ontology with 2854 classes, against a target ontology with 6555 classes. ASMOV found a total of 3347 mappings, with a large number of source ontology classes mapped to multiple target ontology classes.
2.5 Instance Matching
The application of ASMOV to the new set of IIMB instance matching tests results in precision of 86%, recall of 82%, and F1 measure of 84% for the small test, and precision of 85%, recall of 80%, and F1 measure of 82% for the large test.
The results of running the persons and restaurants (PR) tests in the SEALS platform are shown in Table 3. We noted the following issues:
Some conflicts exist between URIs in the TBox (the description ontology) and the ABox. For example, the namespace URI for ontology_people1.owl in the person1 test was http://www.okkam.org/ontology_people1.owl in the TBox but http://www.okkam.org/ontology_person1.owl in the ABox. We manually corrected the TBox file where these differences were found to enable retrieval of the descriptions of the classes and properties used in the ABox.The gold standards for these tests only contained instances of the class “Person” in the person1 and person2 tests, and of the class “Restaurant” in the restaurant test. Running ASMOV in standard fashion produces alignments of instances of other classes such as “Address”; therefore, we restricted ASMOV to only find alignments of a pre-specified class in each test.The gold standard also contains mappings between instances that only match in one specific property, when other potential mappings contain matches in more properties. For example, in the Restaurant test, some matches in the gold standard are done exclusively over the “name” property, even if addresses and other property values are different.The gold standard also contains mappings between instances that have different values for functional properties. For example, the “surname” property is declared as functional for the class “Person” in the TBox, but two instances with “surname” property “carter” and “carcer” respectively are aligned in the gold
131
standard. The semantics of functional properties do not allow such an alignment, and ASMOV therefore rejects it. To test the effect of this, we ran ASMOV against the PR tests using and ignoring the TBox in the description ontology. As can be seen in Table 3,ASMOV obtains better results by ignoring the TBox in both Person tests.
3 General Comments
3.1 Comments on the Results
The current version of ASMOV has shown improvement overall in recall and F1 measure with respect to the results obtained last year in the benchmark tests. This is significant since the results in 2009 were already very high. The larger improvements have been obtained in the most difficult tests, showing the utility of the OAEI benchmarks in driving improvement of alignment algorithms. We have also been able to improve our accuracy in the benchmark, directory, and conference tasks. In the instance matching task we find some differences of interpretation with respect to the gold standard, specifically in terms of the semantics of certain properties.
3.2 Discussions on the Way to Improve ASMOV
ASMOV still needs to improve its ability to work with very large ontologies and resources. While some disk-based storage of partial results has been implemented, the entire contents of the ontologies still needs to loaded in memory prior to performing the matching process. This needs to be further improved to use permanent storage in order to enable the alignment of very large ontologies. We also need to continue the implementation of the ability to infer assertions in order to utilize them for similarity measurement and semantic verification. In addition, we are also working in the improvement of the general scalability of the ASMOV algorithm for the processing of ontologies with a large number of entities. Finally, we need to reexamine the use of an appropriate threshold value to optimize accuracy.
3.3 Comments on the OAEI 2010 Test Cases
The new tests added to the OAEI 2010 contests provide important and welcome tools for the improvement of ontology matching systems. Most importantly, the instance matching task has been made significantly more challenging, allowing us to further refine and expand ASMOV to handle such alignments. Moreover, the availability of
Table 3. PR Instance Matching Results
using TBox ignoring TBoxPerson1 1.000 0.766 1.000 1.000Person2 0.982 0.135 0.701 0.235Restaurants 0.696 0.696 0.696 0.696
132
an ontology in OWL 2 has allowed us to test some of the improvements made to ASMOV in light of the new standard. In addition, the ability to check accuracy using the SEALS system promises to help significantly in the debugging of our algorithms, once the technical problems with SEALS are resolved. Finally, the continuity in the benchmark, anatomy, and conference tracks allows us to evaluate the improvement of our algorithm and implementation as we proceed through its development.
One significant problem we found was the extended downtime encountered with the SEALS system. While it is understandable that some technical issues would be encountered, since this is its first deployment for OAEI, we found that SEALS hindered rather than helped in the process of debugging our algorithm and preparing our results. We trust and expect that many of these problems be resolved in the future, as SEALS promises to be a very useful tool for algorithm evaluation.
4 Conclusion
We have presented a brief description of an automated alignment tool named ASMOV, analyzed its performance at the 2010 Ontology Alignment Evaluation Initiative campaign, and compared it with its 2009 version. The test results show that ASMOV is effective in the ontology alignment realm, and because of its versatility, it performs well in multiple ontology domains such as bibliographic references (benchmark tests) and the biomedical domain (anatomy test). The tests results also showed that ASMOV is a practical tool for real-world applications that require on-the-fly alignments of ontologies.
Acknowledgments. This work is funded by the National Institutes of Health (NIH) under grant R43RR018667. The authors also wish to acknowledge the contribution of Dayana Paez, Michael Ryan, Ray Bradley, and Thomas Taylor of INFOTECH Soft, Inc.
References
1. Euzenat J and Shvaiko P. Ontology Matching. Springer-Verlag, Berlin Heidelberg, 2007.2. Euzenat J. and Valtchev P. Similarity-based ontology alignment in OWL-lite. In Proc. 15th
ECAI, Valencia (ES), 2004, 333-337.3. Euzenat J, et.al. Results of the Ontology Alignment Evaluation Initiative 2009.
http://www.dit.unitn.it/~p2p/OM-2009/oaei09_paper0.pdf. Accessed 1 Oct 2010.4. Jean-Mary Y., Shironoshita E.P., Kabuka, M. Ontology Matching with Semantic
Verification. Journal of Web Semantics. http://dx.doi.org/10.1016/j.websem.2009.04.001.5. Jena from HP Labs http://www.hpl.hp.com/semweb/6. Unified Medical Language System (UMLS) http://umlsks.nlm.nih.gov/7. WordNet http://wordnet.princeton.edu/8. Jean-Mary Y, Kabuka M. ASMOV: Results for OAEI 2009.
http://www.dit.unitn.it/~p2p/OM-2009/oaei09_paper4.pdf.
133
BLOOMS on AgreementMaker: results for OAEI 2010
Catia Pesquita1, Cosmin Stroe 2, Isabel F. Cruz2 , Francisco M. Couto
1
1 Faculdade de Ciencias da Universidade de Lisboa, Portugal
cpesquita-at-xldb.di.fc.ul.pt, fcouto-at-di.fc.ul.pt2 ADVIS Lab, Department of Computer Science, University of Illinois at Chicago
[email protected], [email protected]
Abstract. BLOOMS is an ontology matching method developed as part of an ontology
extension system for biomedical ontologies. It combines two lexical similarity measures with
similarity propagation. These matchers are applied sequentially, following their precision yield:
first lexical similarity based on exact matches, followed by partial matches, and finally these
similarities are propagated throughout the ontologies. Partial matches are based on the
specificity of words within the ontologies vocabularies. Semantic propagation of similarities is
made according to the semantic distance between ontology concepts given by semantic
similarity measures. Alignments are extracted after each matcher, to favor precision, since
BLOOMS was specifically designed to be as automated as possible. For the participation in
OAEI 2010 BLOOMS was integrated into the AgreementMaker system, which provided
ontology loading and navigation capabilities. We participated only in the anatomy track, in the
tasks #1 and #2 (f-measure and precision), given that BLOOMS was specifically designed for
the automated matching of biomedical ontologies. We obtained encouraging results with an f-
measure of 0.828 for task #1 and a precision of 0.967 for task #2. Although the current
implementation of BLOOMS results in very good precision values, recall is below that of the
highest performing systems. This motivates our future work in improving our semantic
propagation algorithm and exploiting external resources.
1 Presentation of the system
BLOOMS is an ontology matching method specifically intended for application to
biomedical ontologies. The matching of biomedical ontologies has become a focus of
interest in recent years due to the increasingly important role that biomedical
ontologies are playing in the knowledge revolution that has swept the Life Sciences
domain in the last decade� ���� ������ ���� � ��� � ���� � ������� � ������ � � � ������������� �������������������������� ���������������� ������������������������������ ��������������������������������� ������������������������������� ���� �� ��� �������������������������� ������������������������������� ��������������������������������������� ����������������Biomedical ontologies present specific challenges and opportunities for their
alignment. One relevant feature of many biomedical ontologies that hinders their
alignment is their size, for instance the Gene Ontology contains over 30,000 concepts
and ChEBI over 500,000. Many of the systems developed for other domains have
134
difficulty in handling such large ontologies. On the other hand, most biomedical
ontologies support few types of relationships, which can hinder the performance of
matchers that explore more complex structures. Also, in most biomedical ontologies
edges do not all represent the same semantic distance between concepts, for instance,
edges deeper in the ontology usually represent shorter distances than edges closer to
the root concept.
Another relevant feature is the rich textual information in the form of concept names,
synonyms and definitions that most biomedical ontologies have. This can play a
crucial role in matching algorithms that exploit lexical resources but it can also be an
obstacle since biomedical terminology has a high degree of ambiguity.
In recent years OAEI has been the major play field for biomedical ontologies
alignment, in its anatomy track. One important finding of previous OAEI anatomy
tracks is that several matches are rather trivial and can be found by simple string
comparison techniques. Based on this notion, the work in [1] has applied a simple
string matching algorithm to several ontologies available in the NCBO BioPortal, and
reported high levels of precision in most cases. There are several possible
explanations for this, including the simple structure of most biomedical ontologies,
their high number of synonyms and low language variability. To improve on the
results of simple string matching, the most successful systems in previous OAEI
editions [2,3] have shown the advantages of two distinct strategies: (1) exploitation of
external knowledge and (2) composition of different matchers followed by
propagation of similarity. The first strategy uses background knowledge resources
such as the UMLS to support lexical matching of concepts [4-6]. The second strategy
propagates similarities between ontology concepts throughout the ontology graphs,
based on the assumption that a match between two concepts should contribute to the
match of their adjacent concepts, according to a propagation factor [7].
BLOOMS was designed to leverage on the success of simple lexical matching
methods, while still finding alignments where lexical similarity is low, by using
global computation techniques. It couples a lexical matching algorithm based on the
specificity of words in the ontology vocabulary, with a novel global similarity
computation approach that takes into account the semantic variability of edges.
1.1 State, purpose, general statement
The original purpose of BLOOMS is to provide the ontology matching component of
an ontology extension system called Auxesia. This system combines ontology
matching and ontology learning techniques to propose new concepts and relations to
biomedical ontologies. Consequently, BLOOMS was specifically designed to match
biomedical ontologies in a fully automated fashion, favoring precision over recall.
Although BLOOMS was specifically designed to be applied to biomedical ontologies,
its current implementation is domain-independent since it can function without
external forms of knowledge. To capitalize on the specific characteristics of most
biomedical ontologies, BLOOMS joins a lexical matcher to exploit the rich textual
component with a global similarity computation technique to handle the cases where
135
synonyms exist but are not shared between ontologies. Furthermore, BLOOMS can
also exploit annotation corpora, which are available for some biomedical ontologies,
to improve the propagation of similarity.
1. Specific techniques used
BLOOMS has a sequential architecture composed of three distinct matchers: Exact,
Partial and Semantic Broadcast Match. While the first two matchers are based on
lexical similarity, the final one is based on the propagation of previously calculated
similarities throughout the ontology graph. Figure 1 depicts the general structure of
BLOOMS.
Figure 1. Diagram of BLOOMS architecture. Given two ontologies, BLOOMS first
extracts alignments based on Exact matches, then on Partial matches, and finally it
propagates the similarities generated by those two strategies using the Semantic
Broadcast approach.
1.2 .1 Lexical similarity
Exact and Partial matchers use lexical similarity based on textual descriptions of
ontology concepts. Textual descriptors of concepts include their labels, synonyms and
definitions. Since ontology concepts usually have several textual descriptors (e.g.,
name, synonyms, definitions), the similarity between two ontology concepts is given
by the maximum similarity between all possible combinations of descriptors.
The first matcher, Exact Match, is run on textual descriptions after normalization and
corresponds to a simple exact match, where the score is either 1 or 0.
136
The second matcher, Partial Match, is applied after processing all concept's labels,
synonyms and definitions through tokenizing strings into words, removing stopwords,
performing normalization of diacritics and special characters, and finally stemming
(Snowball). If the concepts share some of the words in their descriptors, i.e. are partial
matches, the final score is given by a Jaccard similarity, which is calculated by the
number of words shared by the two concepts, over the number of words they both
have. Alternatively, each word can be weighted by its evidence content.
The notion of evidence content (EC) of a word [1] is based on information theory and
can be considered a term relevance measure, since it measures the relevance of a
word within the vocabulary of an ontology. It is calculated as the negative logarithm
of the relative frequency of a word in the ontology vocabulary:
�� ����� ������ �� ������� �������� �
The ontology vocabulary corresponds to all words in all descriptors of all concepts in
the ontology. The final frequency of a word within an ontology corresponds to the
number of concepts that contain it in any of their descriptors. This means that a word
that appears multiple times in the label, definition or synonyms of a concept is only
counted once, preventing bias towards concepts that have many synonyms with very
similar word sets. The evidence content of words that are common to both ontologies,
is given by the average of their ECs within each ontology.
1.2 .2 Semantic Broadcast
After the lexical similarities are computed, they are used as input for a global
similarity computation technique, Semantic Broadcast (SB). This novel approach
takes into account that the edges in the ontology graph do not all convey the same
semantic distance between concepts.
This strategy is based on the notion that concepts whose relatives are similar should
also be similar. A relative of a concept is an ancestor or a descendant whose distance
to the concept is smaller than a factor d. To the initial similarity between concepts, SB
adds the sum of all similarities of the alignments between all relatives weighted by
their semantic gap sG, to a maximum contribution of a factor c. This is given by the
following:
�� �� �� ��� ��������� ��� ������ �� ���� � �� �� � � � � ��� �� � ��� �� � � ��� �� � ��� �� � � � ��� ���� � �� ���
where ca and cb are concepts from ontologies a and b, and ri and rj are relatives of ca
and cb at a distance D smaller than a factor d whose match belongs to the set of
extracted alignments A.
137
The semantic gap between two matches corresponds to the inverse of the average
semantic similarity between the two concepts from each ontology. Several metrics
can be used to calculate the similarity between ontology concepts, in particular,
measures based on information content have been shown to be successful [2].
In BLOOMS we currently implement three information content based similarity
measures: Resnik [3], Lin [4] and a simple semantic difference between each
concept's ICs. The information content of an ontology concept is a measure of its
specificity in a given corpus. Many biomedical ontologies possess annotation corpora
that are suited to this application. Nevertheless, semantic similarity can also be given
by simpler methods based on edge distance and depth.
Semantic broadcast can also be applied iteratively, with a new run using the similarity
matrix provided by the previous.
1.2.3 Alignment Extraction
Alignment extraction in BLOOMS is sequential. After each matcher is run,
alignments are extracted according to a predefined threshold of similarity and
cardinality of matches, so that the concepts already aligned are not processed for
matchers down the line. Each successive matcher has its own predefined threshold.
1.3 Adaptations made for the evaluation
With the purpose of participating in OAEI, BLOOMS was integrated into the
AgreementMaker system [5] due to its extensible and modular architecture. We were
particularly interested in benefiting from its ontology loading and navigation
capabilities, and its layered architecture that allows for serial composition since our
approach combines two matching methods that need to be applied sequentially.
Furthermore, we also exploited the visual interface during the optimization process of
our matching strategy, since although it is not a requirement for our methods, we
found it to be extremely useful, it supports a very quick and intuitive evaluation.
Since neither the mouse or the human anatomy ontologies have an annotation corpus,
the Semantic Broadcast algorithm used a semantic similarity measure based on edge
distance and depth, where similarity decreases with the number of edges between two
concepts, and edges further away from the root correspond to higher levels of
similarity.
2 Results
BLOOMS was only submitted to the anatomy track, since it is being specifically
developed to handle biomedical ontologies. The anatomy track contains 4 tasks: in the
first three tasks, matchers should be optimized to favor f-measure, precision and
recall, in turn. In the fourth task, an initial set of alignments is given, that can be used
138
to improve the matchers performance. In addition to the classical measures of
precision, recall and f-measure, the OAEI initiative also employs recall+, which
measures the recall of non-trivial matches, since in the anatomy track a large
proportion of matches can be achieved suing simple string matching techniques.
We only participated in tasks #1 and #2, since BLOOMS is designed to favor
precision.
2.1 anatomy
Taking advantage of the SEALS platform we ran several distinct configurations of
BLOOMS, testing different parameters and also analyzing the contribution of each
matcher to the final alignment.
We found that after the first matcher is run, the alignments produced have a very high
precision (0.98), but the recall is somewhat low (0.63). Each of the following
matchers increases recall while slightly decreasing precision, which was expected
given the increasing laxity they provide.
We also found that weighting the partial match score using word evidence content did
not significantly alter results when compared to the simple Jaccard similarity.
For task #1 we used a Partial Match threshold of 0.9 and a final threshold of 0.4.
Semantic Broadcast was run to propagate similarities through ancestors and
descendants at a maximum distance of 2, and contribution was set to 0.4. Using the
SEALS evaluation platform, we obtained 0.954 precision, 0.731 recall, for a final F-
measure of 0.828 and a recall+ of 0.315.
For task #2 we used a Partial Match threshold of 0.9 and did not use Semantic
Broadcast. With this strategy, we ensured a higher precision, of 0.967. However,
recall was not much lower than the one in task #1, 0.725 , which resulted in a final f-
measure of 0.829.
We did not participate in other tasks, since BLOOMS was originally intended to yield
a high precision, as it is intended to be run in a fully automated fashion as a part of an
ontology extension system.
3 General comments
We find that the SEALS platform is a very valuable tool in improving matching
strategies. We find however that the 100 minute time limit might be detrimental to
strategies that need to process large external resources.
3.1 Comments on the results
BLOOMS was designed to be as fully automated as possible, so it is more geared
towards increased precision than recall. Comparing our results for tasks #1 and #2,
they clearly indicate that our semantic broadcast strategy does not represent a very
139
heavy contribution to recall, but that we do capture nearly 10% more matches when
using both the Exact and Partial Match strategies, than Exact Match alone. Also our
recall+ is not very high, again highlighting the need to expand our strategy to improve
recall.
Nevertheless, we find our performance to be comparable to the best systems in 2009,
and in 2010 our f-measure in task #1 is 5% lower that the best performing system,
whereas in task #2 we are the second best system, with a slight difference of 0.1% in
precision. These are encouraging results and we fully intend to participate in future
events with an improved version of BLOOMS.
3.2 Discussions on the way to improve the proposed system
We are planning on implementing several strategies for improvement in the near
future, some of which were already a part of our initial strategy, but were not yet
implemented at the time of OAEI 2010. To improve the lexical similarity matchers,
future versions of BLOOMS will take into account spelling variants and mistakes,
and we will also investigate the feasibility of using external resources such as UMLS
and WordNet to increase the number of synonyms for both terms and words. We feel
this would greatly improve the recall of our strategy. Regarding similarity
propagation, we will work extensively on improving our semantic broadcast
approach, by exploring alternative strategies for the computation of information
content independently of an annotation corpus, and thus expand the number of
semantic similarity measures that can be used. We will also adapt semantic broadcast
to propagate dissimilarity, and decrease the similarity between concepts that might
have a high lexical similarity but very distinct neighborhoods.
4 Conclusion
Participating in the anatomy track of OAEI 2010 has given us an opportunity to
evaluate a matching algorithm developed with the practical purpose of being used in a
semi-automated ontology extension system, Auxesia. Our matching algorithm,
BLOOMS, is intended to be as automated as possible, and thus its current
implementation favors precision. This was clearly visible in the results we obtained in
tasks #1 and #2 of the anatomy track of OAEI 2010, where we obtained high ranking
precision values within the top 3, but lower recall.
In future versions of BLOOMS we will implement several strategies designed to
improve recall, while minimizing precision loss.
The lessons learned throughout this period will undoubtedly contribute to an
improvement of our method.
140
Acknowledgements
The work performed at University of Lisbon by Catia Pesquita and Francisco M.
Couto was supported by the Multiannual Funding Program and the PhD grant SFRH/
BD/42481/2007.
The work performed at UIC by Cosmin Stroe and Isabel F. Cruz has been
partially sponsored by the National Science Foundation under Awards
IIS-0513553 and IIS-0812258
References
1. Ghazvinian, A., Noy, N. , Musen, M. (2009). Creating mappings for ontologies in
biomedicine: Simple methods work. In AMIA Annual Symposium (AMIA 2009), 2.
73
2. Caracciolo, C., Hollink, L., Ichise, R., Meilicke, C., Pane, J. , Shvaiko, P. (2008).
Results of the Ontology Alignment Evaluation Initiative 2008. The 7th International
Semantic Web Conference. 33, 73
3. Ferrara, A., Hollink, L., Isaac, A., Joslyn, C., Meilicke, C., Nikolov, A., Pane, J.,
Shvaiko, P., Spiliopoulos, V. , Wang, S. (2009). Results of the Ontology Alignment
Evaluation Initiative 2009. Fourth International Workshop on Ontology Matching,
Washington, DC , 1. 16, 33, 74
4. Zhang, S., Bodenreider, O. (2007). Lessons Learned from Cross-Validating
Alignments between Large Anatomical Ontologies. MedInfo, 12, 822--826. 33
5. Lambrix, P., Tan, H. (2006). SAMBO - system for aligning and merging
biomedical ontologies. Web Semantics: Science, Services and Agents on the World
Wide Web, 4, 196--206. 31, 33
6. Jean-Mary, Y.R., Shironoshita, E.P. , Kabuka, M.R. (2009). Ontology matching
with semantic verification. Web Semantics: Science, Services and Agents on the
World Wide Web, 7, 235--251. 33
7. Cruz, I.F., Antonelli, F.P., Stroe, C. (2009). AgreementMaker: Efficient Matching
for Large Real-World Schemas and Ontologies. PVLDB, 2, 1586- 1589. 82
8. Couto, F., Silva, M. & Coutinho, P. (2005). Finding genomic ontology terms in text
using evidence content. BMC Bioinformatics, 6, S21. 57, 64
9. Pesquita, C., Faria, D., Falcao, A.O., Lord, P., Couto, F.M., Bourne, P.E. (2009).�
Semantic Similarity in Biomedical Ontologies. PLoS Computational Biology, 5
10. Resnik, P. (1998). Semantic Similarity in a Taxonomy: An Information-Based
Measure and its Application to Problems of Ambiguity in Natural Language. Journal
of Artificial Intelligence Research.
11. Lin D (1998) An information-theoretic definition of similarity. Proc. of the 15th
International Conference on Machine Learning. San Francisco, CA: Morgan
Kaufmann. pp. 296–304.
141
CODI: Combinatorial Optimization for DataIntegration – Results for OAEI 2010
Jan Noessner and Mathias Niepert
KR & KM Research GroupUniversity of Mannheim, Germany
{jan, mathias}@informatik.uni-mannheim.de
Abstract. The problem of linking entities in heterogeneous and decentralizeddata repositories is the driving force behind the data and knowledge integrationeffort. In this paper, we describe our probabilistic-logical alignment system CODI(Combinatorial Optimization for Data Integration). The system provides a declar-ative framework for the alignment of individuals, concepts, and properties of twoheterogeneous ontologies. CODI leverages both logical schema information andlexical similarity measures with a well-defined semantics for A-Box and T-Boxmatching. The alignments are computed by solving corresponding combinatorialoptimization problems.
1 Presentation of the system
1.1 State, purpose, general statement
CODI (Combinatorial Optimization for Data Integration) leverages terminological struc-ture for ontology matching. The current implementation produces mappings betweenconcepts, properties, and individuals including mappings between object and data typeproperties. The system combines lexical similarity measures with schema informationto reduce or completely avoid incoherence and inconsistency during the alignment pro-cess. The system is based on the syntax and semantics of Markov logic [2] and trans-forms the alignment problem to a maximum-a-posteriori optimization problem.
1.2 Specific techniques used
Markov logic combines first-order logic and undirected probabilistic graphical mod-els [11]. A Markov logic network (MLN) is a set of first-order formulae with weights.Intuitively, the more evidence there is that a formula is true the higher the weight ofthis formula. It has been proposed as a possible approach to several problems occur-ring in the context of the semantic web [2]. We have shown that Markov logic providesa suitable framework for ontology matching as it captures both hard logical axiomsand soft uncertain statements about potential correspondences between entities. Theprobabilistic-logical framework we propose for ontology matching essentially adaptsthe syntax and semantics of Markov logic. However, we always type predicates andwe require a strict distinction between hard and soft formulae as well as hidden andobservable predicates. Given a set of constants (the classes and object properties of
142
the ontologies) and formulae (the axioms holding between the objects and classes), aMarkov logic network defines a probability distribution over possible alignments. Werefer the reader to [9, 8] for an in-depth discussion of the approach and some compu-tational challenges. For generating the Marcov logic networks we used the approachdescribed in [12].
T-Box Matching Formalization Given two ontologies O1 and O2 and an initial a-priori similarity measure σ we apply the following formalization. First, we introduceobservable predicates O to model the structure of O1 and O2 with respect to both con-cepts and properties. For the sake of simplicity we use uppercase letters D,E,R torefer to individual concepts and properties in the ontologies and lowercase letters d, e, rto refer to the corresponding constants in C. In particular, we add ground atoms ofobservable predicates to Fh for i ∈ {1, 2} according to the following rules1:
Oi |= D � E �→ subi(d, e)
Oi |= D � ¬E �→ disi(d, e)
Oi |= ∃R.� � D �→ subdi (r, d)
Oi |= ∃R.� � D �→ supdi (r, d)
Oi |= ∃R.� � ¬D �→ disdi (r, d)
The ground atoms of observable predicates are added to the set of hard constraints Fh,forcing them to hold in computed alignments. The hidden predicates mc and mp, on theother hand, model the sought-after concept and property correspondences, respectively.Given the state of the observable predicates, we are interested in determining the stateof the hidden predicates that maximize the a-posteriori probability of the correspondingpossible world. The ground atoms of these hidden predicates are assigned the weightsspecified by the a-priori similarity σ. The higher this value for a correspondence themore likely the correspondence is correct a-priori. Hence, the following ground formu-lae are added to Fs:
(mc(c, d), σ(C,D)) if C and D are concepts
(mp(p, r), σ(P,R)) if P and R are properties
Notice that the distinction between mc and mp is required since we use typed predicatesand distinguish between the concept and property type.
Cardinality Constraints A method often applied in real-world scenarios is the se-lection of a functional one-to-one alignment [1]. Within the ML framework, we caninclude a set of hard cardinality constraints, restricting the alignment to be functionaland one-to-one. In the following we write x, y, z to refer to variables ranging over theappropriately typed constants and omit the universal quantifiers.
mc(x, y) ∧mc(x, z) ⇒ y = z
mc(x, y) ∧mc(z, y) ⇒ x = z
Analogously, the same formulae can be included with hidden predicates mp, restrictingthe property alignment to be one-to-one and functional.
1 Due to space considerations the list is incomplete. For instance, predicates modeling rangerestrictions are not included.
143
Coherence Constraints Incoherence occurs when axioms in ontologies lead to log-ical contradictions. Clearly, it is desirable to avoid incoherence during the alignmentprocess. All existing approaches to alignment repair remove correspondences after thecomputation of the alignment. Within the ML framework we can incorporate incoher-ence reducing constraints during the alignment process for the first time. This is accom-plished by adding formulae of the following type to Fh.
dis1(x, x′) ∧ sub2(x, x
′) ⇒ ¬(mc(x, y) ∧mc(x′
, y′))
disd1(x, x
′) ∧ subd2(y, y
′) ⇒ ¬(mp(x, y) ∧mc(x′
, y′))
Stability Constraints Several approaches to schema and ontology matching propa-gate alignment evidence derived from structural relationships between concepts andproperties. These methods leverage the fact that existing evidence for the equivalenceof concepts C and D also makes it more likely that, for example, child concepts of Cand child concepts of D are equivalent. One such approach to evidence propagation issimilarity flooding [7]. As a reciprocal idea, the general notion of stability was intro-duced, expressing that an alignment should not introduce new structural knowledge [5].The soft formula below, for instance, decreases the probability of alignments that mapconcepts X to Y and X ′ to Y ′ if X ′ subsumes X but Y ′ does not subsume Y .
(sub1(x, x′) ∧ ¬sub2(y, y
′) ⇒ mc(x, y) ∧mc(x′
, y′), w1)
(subd1(x, x′) ∧ ¬subd2(y, y
′) ⇒ mp(x, y) ∧mc(x′
, y′), w2)
Here, w1 and w2 are negative real-valued weights, rendering alignments that satisfy theformulae possible but less likely.
The presented list of cardinality, coherence, and stability constraints could be ex-tended by additional soft and hard formulae. Other constraints could, for example,model known correct correspondences or generalize the one-to-one alignment to m-to-n alignments.
A-Box Matching The current instance matching configuration of CODI leverages ter-minological structure and combines it with lexical similarity measures. The approachis presented in more detail in [10]. It uses one T-Box T but two different A-BoxesA1 ∈ O1 and A2 ∈ O2. In cases with two different T-Boxes the T-Box matching ap-proach is applied as a preprocessing step, merge the two aligned T-Boxes and then useour instance matching algorithm. CODI offers complete conflict elimination meaningthat the resulting alignment is always coherent for OWL DL ontologies. This compo-nent is based on the work of Meilicke et al. [6]. CODI enforces the instance alignmentto be consistent. To this end, we need to introduce observable predicates O to modelconflicts, that is, a positive assertion of one instance in one ontology and a negativeassertion of the same instance in the other ontology. This is done for both property andconcept assertions.
Analogous to the concept and property alignment before, we introduce the hiddenpredicate mi representing instance correspondences. Let C be a concept and P be aproperty of T-Box T . Further, let A ∈ A1 and B ∈ A2 be individuals in the respectiveA-Boxes. Then, using a reasoner, ground atoms are added to the set of hard constraints
144
Fh according to the following rules:
T ∪ A1 |= C(A) ∧ T ∪ A2 |= ¬C(B) �→ ¬mi(a, b)
T ∪ A1 |= ¬C(A) ∧ T ∪ A2 |= C(B) �→ ¬mi(a, b)
T ∪ A1 |= P (A,A′) ∧ T ∪ A2 |= ¬P (B,B
′) �→ ¬mi(a, b) ∨ ¬mi(a′
, b′)
T ∪ A1 |= ¬P (A,A′) ∧ T ∪ A2 |= P (B,B
′) �→ ¬mi(a, b) ∨ ¬mi(a′
, b′)
In addition to these formulae we included cardinality constraints analogous to thoseused in the concept and property matching of Section 1.2. In the instance matching for-mulation, the a-priori similarity σc and σp measures the normalized overlap of conceptand property assertions, respectively. For more details on these measures, we refer thereader to [10]. The following formulae are added to the set of soft formulae Fs:
(mi(a, b), σc(A,B)) if A and B are instances
(mi(a, b) ∧mi(c, d), σp(A,B,C,D)) if A, B, C, and D are instances
1.3 Adaptations made for the evaluation
The strength of the system is its modularity allowing the incorporation of different simi-larity measures. The system can be optimized in two major ways: (a) Inclusion of novelformulae enforcing the logical consistency and (b) the inclusion of additional similaritymeasures. There is room for improvement since we used a very simple lexical similar-ity measure based on the Levenshtein distance [4] for our experiments. It is possible toapply different aggregation functions like average or maximum and to include specificproperties of an ontology like URIs, labels, and comments.
In all OAEI test cases Algorithm 1 was used for computing the a-priori similarityσ(entity1, entity2). In the case of concept and property alignments, the a-priori simi-larity is computed by taking the maximal similarity between the URIs, labels and OBOto OWL constructs. In case of instance matching the algorithm goes through all dataproperties and takes the average of the similarity scores.
1.4 Link to the System and Parameters File
CODI can be downloaded from http://codi-matcher.googlecode.com.
1.5 Link to the Set of Provided Alignments
The alignments for the tracks Benchmark and Conference has been made with theSEALS platform. For Anatomy, IIMB, and Restaurant the alignments can be foundat http://code.google.com/p/codi-matcher/downloads/list
2 Results
In the following section, we present the results of the CODI system for the individualOAEI tracks. Due to space considerations, we do not explain the different benchmarksin more detail.
145
Algorithm 1 σ(entity1, entity2)
if entity1 and entity2 are either concepts or properties thenvalue ← 0for all Values s1 of URI, labels, and OBOtoOWL constructs in entity1 do
for all Values s2 of URI, labels, and OBOtoOWL constructs in entity1 dovalue ← Max(value, sim(s1, s2))
end forend forreturn value
end ifif entity1 and entity2 are individuals then
Map〈URI, double〉 similarities ← null
for all dataproperties dp1 of entity1 douri1 ← URI of dp1for all dataproperties dp2 of entity2 do
if uri1 equals URI of dp2 thenvalue ← sim(valueofdp1, valueofdp2)if uri1 is entailed in similarities then
update entry 〈uri1, old value〉 to 〈uri1, Minimum (old value+ value, 1)〉 insimilarities
elseadd new entry pair 〈uri1, value〉 in similarities
end ifend if
end forend forreturn (sum of all values in similarities)/(length of similarities)
end if
146
Benchmark Track While our system’s strength is its modularity and adaptability todifferent ontologies we used the exact same setting for all ontology matching tracks.Hence, the performance on the benchmark track is rather poor. This is primarily dueto the high threshold of 0.85 for the Levenshtein similarity measure that we applied ineach of the ontology matching tracks. The results are shown in Table 1.
Table 1. Benchmark results
1xx 2xx 3xx AveragePrecision 1 0.70 0.92 0.72Recall 0.99 0.42 0.43 0.44F1 score 1 0.49 0.56 0.51
Conference Track On the real-world conference dataset CODI achieves very goodresults since it employs logical reasoning to avoid incoherences. The execution time isbetween 2 and 4 minutes per test case2. Table 2 summarizes the overall results.
Table 2. Conference results
AveragePrecision 0.87Recall 0.51F1 score 0.64
Anatomy Track The results on the anatomy track are also convincing. The resultsshown in Table 3 are en par with the 2009 results of state-of-the-art matching applica-tions. The F1 scores are between 0.79 and 0.73 for all subtasks, even for the two tasksFocus on Precision and Focus on Recall. Thus, our algorithm achieves satisfiable pre-cision and recall values without sacrifices on the F1 score. For the last task, where apartial reference alignment was given, we could gain almost 5 % on the F1 score. Thisis because incorporating a partial reference alignment in our system is straight-forward.The reference alignment becomes a direct part of the optimization problem, enforcinggood correspondences while ruling out contradicting ones. However, since our algo-rithm uses logical reasoning and has to solve an NP-hard optimization problem, theexecution times are quite high3.
Table 3. Anatomy results
Focus on Focus on Focus on PartialF1 score Precision Recall Alignment
Precision 0.954 0.964 0.782 0.969Recall 0.680 0.663 0.695 0.742F1 score 0.794 0.784 0.736 0.840Execution Time (min) 88 60 157 95
2 All experiments are executed on a Desktop PC with 2 GB RAM and a Intel Core2 Duo 2.4GHz processor.
3 This forces us to submit the solutions without the seals platform because of a timeout after 45minutes.
147
IIMB Track The instance matching benchmark IIMB consists of 80 transformationsdivided in four transformation categories containing 20 transformations each. We ap-plied the full A-Box matching functionality described above with a threshold on thea-priori similarity of 0.1. The average execution time on the IIMB small (large) datasetis 2.6 (35.1) minutes. Table 4 summarizes the different results of the CODI system.The values without brackets are the results for the small IIMB dataset and the values inbrackets for the large one.
Table 4. IIMB results
Transformations 0-20 21-40 41-60 61-804 overallPrecision 0.99 (0.98) 0.95 (0.94) 0.96 (0.99) 0.86 (0.86) 0.94 (0.95)Recall 0.93 (0.87) 0.83 (0.79) 0.97 (0.99) 0.54 (0.53) 0.83 (0.80)F1 score 0.96 (0.91) 0.88 (0.85) 0.97 (0.99) 0.65 (0.63) 0.87 (0.85)
PR Track For this track consisting of small files about persons and restaurants, weused a simple one to one alignment only based on lexical similarity scores since nosignificant structural information is available. Thus, the runtime was with less than 5seconds per test case very short. The results of the CODI system are depicted in Table 5.
Table 5. PR results
Person1 Person2 RestaurantPrecision 0.87 0.83 0.71Recall 0.96 0.22 0.72F1-score 0.91 0.36 0.72
3 General comments
3.1 Discussions on the way to improve the proposed system
CODI is a very young system and does not yet provide a user interface. Hence, im-provements in usability by designing a suitable user interface will be one of the nextsteps. In case of the quality of the alignments, more sophisticated lexical similarity mea-sures will be tested and integrated. We are also working on novel algorithms solving theoptimization problems more efficiently.
3.2 Comments on the OAEI 2010 procedure
The SEALS evaluation campaign is very beneficial since it is the first time that thematchers must have a standardized interface which could possibly be used by everyone.
3.3 Comments on the OAEI 2010 measures
We encorage the organizers to use semantic precision and recall measures as describedin [3].
148
4 Conclusion
CODI performs concept, property, and instance alignments. It combines logical andstructural information with a-priori similarity measures in a well-defined way by usingthe syntax and semantics of Markov logic. The system therefore not only aligns the en-tities with the highest lexical similarity but also enforces the coherence and consistencyof the resulting alignment.
The overall results of the young system are very promising. Especially when con-sidering the fact that there are many optimization possibilities with respect to the lexicalsimilarity measures that have not yet been investigated. The strength of the CODI sys-tem is the combination of lexical and structural information and the declarative naturethat allows easy experimentation. We will continue the development of the CODI sys-tem and hope that our approach inspires other researchers to leverage terminologicalstructure for ontology matching.
References
1. I. Cruz, F. Palandri, Antonelli, and C. Stroe. Efficient selection of mappings and automaticquality-driven combination of matching methods. In Proceedings of the ISWC 2009 Work-shop on Ontology Matching, 2009.
2. P. Domingos, D. Lowd, S. Kok, H. Poon, M. Richardson, and P. Singla. Just add weights:Markov logic for the semantic web. In Proceedings of the Workshop on Uncertain Reasoningfor the Semantic Web, pages 1–25, 2008.
3. D. Fleischhacker and H. Stuckenschmidt. A Practical Implementation of Semantic Precisionand Recall. In 2010 International Conference on Complex, Intelligent and Software IntensiveSystems, pages 986–991. IEEE, 2010.
4. V. I. Levenshtein. Binary codes capable of correcting deletions and insertions and reversals.Doklady Akademii Nauk SSSR, pages 845–848, 1965.
5. C. Meilicke and H. Stuckenschmidt. Analyzing mapping extraction approaches. In Proceed-ings of the Workshop on Ontology Matching, Busan, Korea, 2007.
6. C. Meilicke, A. Tamilin, and H. Stuckenschmidt. Repairing ontology mappings. In Pro-ceedings of the Conference on Artificial Intelligence, pages 1408–1413, Vancouver, Canada,2007.
7. S. Melnik, H. Garcia-Molina, and E. Rahm. Similarity flooding: A versatile graph matchingalgorithm and its application to schema matching. In Proceedings of ICDE, pages 117–128,2002.
8. M. Niepert. A Delayed Column Generation Strategy for Exact k-Bounded MAP Inference inMarkov Logic Networks. In Proceedings of the 25th Conference on Uncertainty in ArtificialIntelligence, 2010.
9. M. Niepert, C. Meilicke, and H. Stuckenschmidt. A Probabilistic-Logical Framework forOntology Matching. In Proceedings of the 24th AAAI Conference on Artificial Intelligence,2010.
10. J. Noessner, M. Niepert, C. Meilicke, and H. Stuckenschmidt. Leveraging TerminologicalStructure for Object Reconciliation. The Semantic Web: Research and Applications, pages334–348, 2010.
11. M. Richardson and P. Domingos. Markov logic networks. Machine Learning, 62(1-2):107–136, 2006.
12. S. Riedel. Improving the accuracy and efficiency of map inference for markov logic. In Pro-ceedings of the Conference on Uncertainty in Artificial Intelligence, pages 468–475, 2008.
149
Eff2Match Results for OAEI 2010
Watson Wei Khong Chua1 and Jung-Jae Kim2
1 Nanyang Technological University, Singapore [email protected] Nanyang Technological University, Singapore [email protected]
Abstract. While the primary objective of an ontology alignment tool is to iden-
tify as many correct correspondences as possible, efficiency in terms of run-time
needs to be achieved for practical usage. Not only does run-time efficiency enable
scalability, it also facilitates information integration for time-critical applications
using heterogeneous ontologies. In this paper, we present our ontology alignment
approach known as Eff2Match which aligns a pair of ontologies with high accu-
racy and low runtime.
1 Presentation of the system
Ontologies are being widely used for semantic representation in applications from var-
ious domains such as biomedical informatics [3] and earth sciences [5]. They can be
used to provide data with semantics, thus resolving the heterogeneity problem be-
tween information sources at the data level. However, the problem is only partially
resolved because different ontology engineers model their ontologies differently. There-
fore, the heterogeneity problem is escalated to the ontological level if information is
to be shared between applications using different ontologies. As a result, ontology
alignment tools that can achieve high accuracy are required. In this paper, we present
Eff2Match (pronounced “Eff Squared match”), an Effective and Efficient ontology
matching tool which can match a pair of ontologies with good accuracy within a short
amount of time. Eff2Match uses an effective and dynamic candidate reduction technique
to avoid performing unnecessary comparisons, thereby achieving high efficiency.
1.1 State, purpose, general statement
In order to facilitate the sharing of information among applications using different on-
tologies, we have developed an automatic ontology alignment tool called Eff2Match. It
is able to align both concepts and properties in different ontologies that are semantically
equivalent (concepts are matched to concepts and properties are matched to properties).
The current implementation does not match the instances in the ontologies.
1.2 Specific techniques used
Eff2Match takes as input the URI of a pair of ontologies to be aligned and matches
entities (concepts or properties) in the source ontology to those in the target ontology.
The alignment process consists of four stages: 1) Anchor Generation, 2) Candidates
Generation, 3) Anchor Expansion and 4) Iterative Score Boosting as shown in Fig. 1 .
150
1) AnchorsGeneration
2) CandidateGeneration
3) AnchorsExpansion
4) IterativeBoosting
Initialanchor set,
Ain
Expandedanchor set,
Aexp
CandidateSet
Newmatches
Mnew
FinalAlignment
Afin
Newanchor set,
Anew
U
Remove pairs in AnewO1 O2
Remove pairs in Ain
U
Fig. 1. Eff2Match Algorithm Flow
Anchor Generation In the Anchor Generation stage, matching entities are identified
using an exact string matching technique. Local names and labels of entities e2j in the
target ontology are first preprocessed through camel case conversion, case-normalization
and removal of delimiters. A hash-table is then used to map the preprocessed local
names and labels to their corresponding entities. After that, we preprocess the local
name and label for each entity e1i in the source ontology and look them up in the
hash-table. If either a matching local name or label can be found in the hash-table,
we consider the corresponding entity e2j in the target ontology to be equivalent to the
source entity. This method is significantly faster than a pairwise comparison of local
names and labels as it takes only O(n1) + O(n2) time compared to the latter which
requires O(n1 ×n2) where n1 and n2 are the number of entities in the source ontology
O1 and target ontology O2 respectively.
Candidate Generation In the Candidates Generation stage, we enumerate candidates
for entities in the source ontology that has not been matched in the previous stage using
a Vector Space Model (VSM) approach. For each concept, we generated three VSM
vectors from the annotations (local name, label and comments) in the ancestors (V eca),
descendants (V ecd) and the concept (V ecc) itself. For each property, the vectors gen-
erated consist of the annotations in the property (V ecp) itself, the property’s domain
concepts (V ecdo) and range concepts(V ecr). The VSM similarity (SimV SM ) between
two concepts c1i and c2j is an aggregation of the cosine similarity between the concept
VSM vectors, ancestor VSM vectors and descendant VSM vectors using a weighted
151
average.
SimV SM =α× V ecc1i · V ecc2j + β × V eca2i · V eca2j + γ × V ecd2i · V ecd2j
α+ β + γ
where α, β and γ are the weights given to the similarity between annotations of the
concepts themselves, annotations of their ancestor concepts and annotations of their
descendant concepts respectively. The similarity values for properties are calculated in
a similar manner and two matrices, Mcon and Mprop are used to store the similarity
values for concepts and properties respectively. The VSM similarities are normalised to
[0,1] by dividing each entry in Mcon and Mprop by their largest value to get the nor-
malised VSM similarity, SimV SMN(C1i, C2j). Candidates selection is then performed
for each source entity by taking the top-K entities in the target ontology according to
their VSM similarities.
Anchor Expansion In the anchor expansion stage, more equivalent pairs of entities are
identified by comparing the source entities with their candidate entities using termino-
logical methods. In Eff2Match, a term-removing algorithm (TRA) is used for efficiency
purposes and the algorithm is illustrated in Fig. 2. First, the labels (local names) of a
pair of entities are tokenised. The tokens are then stemmed and words that are stemmed
to the same form are removed from both labels. If there are tokens remaining in both
the labels, the next stage compares tokens from different labels pairwise using WordNet
to determine if they are synonyms of each other. Synonymous tokens are then removed
from the labels. If there are no tokens remaining in both the labels after any stage, the
two entities are considered to be equivalent and added to the anchor set.
If there are remaining tokens in only one of the labels and not the other, we use
a novel technique known as Informative Word Matching (IWM) to determine if the
entities are matches. If concept C1i contains p more terms than C2j and these p terms
occur in the labels of the ancestors of C2j , we can consider C1i and C2j to have the
same meaning. For example, if C1i has the label Heart Endocardium and C2j has the
label Endocardium and the C2j is a sub-concept of Heart Part, we can determine that
C1i and C2j are semantically equivalent from their labels since the word Heart in C1iis not informative.
Given a pair of entities eemp and erem where eemp is the entity without remain-
ing tokens in its label and erem is the entity with p remaining tokens in its label, the
following steps are performed to determine if the p remaining tokens are informative
words:
1. Collect the labels of ancestors of eemp up to r generations or when the root of the
ontology is reached, whichever is earlier.
2. Tokenise and stem the collection of labels collected to get a set of stemmed ancestor
tokens Ssat.3. For each token ti, i ∈ [1..p] remaining in erem, stem ti and check if it exists in
Ssat. If it does, the word is not informative and it is removed it from erem.
4. Look up the definition of the original label of eemp in WordNet, tokenise and stem
the words in the definition before adding them to the set of definition words, Sdef .
152
5. For each token ti, i ∈ [1..p] remaining in erem, stem ti and check if it exists in
Sdef . If it does, the word is not informative and it is removed it from erem.
Add entity pair toexpanded anchor
set
Remove tokensstemmed to the
same form
Remove tokensthat are
synonymous
Checkinformativeness of
excess words
Are both labelsempty?
Tokenise andremove stopwords
Is one of thelabels empty?
Are both labelsempty?
Is one of thelabels empty?
No
No
No
Yes
Yes
Yes
Yes
Are excesswords
informative?
No
Fig. 2. Anchor Expansion Flow Chart
Iterative Boosting In the final stage of the matching process, an iterative boosting (Iter-
Boost) process is used to identify more pairs of equivalent concepts using the expanded
anchor set Aexp. In this stage, the algorithm attempts to match the source concepts that
have not been matched with their candidates iteratively. In each iteration, the source
concepts are ranked based on the sum of their ancestors and descendants with matches
in Aexp. The top K source concepts are then selected and a formula is used to boost the
score of their candidates based on the number of common ancestors and descendants
that they share. Given a source concept C1i and a candidate concept C2j , the structural
overlap SO(C1i, C2j) between them is calculated using:
SO(C1i, C2j) =|ξ(A(C1i), A(C2j))|+ |ξ(D(C1i), D(C2j))|
min(|A(C1i)|, |A(C2j)|) +min(|D(C1i)|, |D(C2j)|)
where A(C) is the set of ancestors of concept C and D(C) is the set of descendants
of concept C and the function ξ(X,Y ) enumerates the set of concepts in X which
153
have equivalences in Y. The equivalences were determined from the comparison in the
previous stages. The similarity between C1i and C2j is then given by:
Simboost(C1i, C2j) =
{SimV SMN (C1i, C2j), SO(C1i, C2j) > tb√SimV SMN
(C1i, C2j), otherwise.
The highest scoring candidate is then selected to be the matching concept and the confi-
dence that the pair matches is given by Simboost(C1i, C2j). If Simboost(C1i, C2j) is
greater a cut-off threshold tc, C1i and C2j are inserted intoAexp as well as the final set
of alignment and the process is repeated until all the source entities and their candidates
have been visited. If Simboost(C1i, C2j) is less than tc, C1i and C2j are inserted into
the set of final alignment but are not considered anchors for future iterations.
1.3 Adaptations made for the evaluation
No special adaptations were made for individual tracks and all alignment processes
make use of the same set of parameters. The cut-off threshold for the correspondences
was set at 0.7 for the best F-Measure. The only external resource that we used is Word-
Net. For Informative Word Matching, we set p = 1, meaning that we only perform IFM
for entities with one remaining token after the Term-Removing Algorithm (TRA). For
IterBoost, the cut-off threshold for boosting, tb is set at 0.4 while the cut-off threshold
for matching entities to be considered anchors, tc is set at 0.5. Lastly, the weights α, β,
and γ are set to 2, 1, 1, respectively. The matcher has also been implemented as a web
service so that it can be evaluated on the SEALS platform.
1.4 Link to the system and parameters file
The Eff2Match system (jar file) and configuration files can be found at 3
http://www.cais.ntu.edu.sg/˜chua0507/OAEI/Eff2MatchSystem.zipwith installation instructions at
http://www.cais.ntu.edu.sg/˜chua0507/OAEI/Instructions.txt
1.5 Link to the set of provided alignments (in align format)
The set of alignments produced by Eff2Match can be found at
http://www.cais.ntu.edu.sg/˜chua0507/OAEI/Eff2MatchAlignments.zip
2 Results
Eff2Match participated in the benchmark, anatomy and conference tracks of the OAEI
2010 competition and the results are presented in the following subsections:
3 Please replace the ˜symbol by typing it in manually using your keyboard.
154
2.1 Benchmark
The ontologies in the benchmark dataset can be categorised into 5 different categories
according to the difficulty of matching as shown in Table 2.1. Eff2Match performs well
on all the categories, achieving an F-Measure of more than 0.75 with the exception of
the category 248− 266. Ontologies in this category have different linguistics and struc-
tural characteristics, which are core attributes which Eff2Match relies on for finding
correspondences, thus explaining the poor results.
Ontologies Precision Recall F-Measure101-104 1.000 1.000 1.000
201-210 0.986 0.684 0.768
221-247 0.990 1.000 0.995
248-266 0.929 0.502 0.591
301-304 0.889 0.711 0.780
Table 1. Results for benchmark dataset
2.2 Anatomy
The anatomy dataset consists of two large real world ontologies, namely the Adult
Mouse Anatomy with 2247 classes and the anatomy part of the NCI Thesaurus with
3304 classes. Eff2Match’s results for this track are shown in Table 2.2 and shows that
Eff2Match can match large ontologies with high accuracy and short run-time.
Task Precision Recall F-Measure Recall+ Time TakenTask 1 (Optimal F-Measure) 0.955 0.781 0.859 0.440 2.5 mins
Task 2 (Optimal Precision) 0.968 0.745 0.842 N.A. 2.5 mins
Task 3 (Optimal Recall) 0.766 0.838 0.800 0.588 2.5 mins
Table 2. Results for anatomy dataset
2.3 Conference
Lastly, Table 2.3 presents results for Eff2Match in the conference track for the 21 on-
tologies where reference alignments are available. As these ontologies are developed
heterogeneously, discovering the correct correspondences between them is more diffi-
cult than for the other two tracks. Eff2Match was able to achieve F-Measures ranging
from 0.4 to 0.759 for pairs of ontologies in this track.
155
Ontologies Precision Recall F-Measurecmt-ekaw 0.316 0.545 0.400
conference-edas 0.303 0.588 0.400
conference-iasted 0.350 0.500 0.412
cmt-iasted 0.267 1.000 0.421
edas-iasted 0.471 0.421 0.444
cmt-conference 0.467 0.438 0.452
cmt-confof 0.538 0.438 0.483
conference-ekaw 0.481 0.520 0.500
edas-ekaw 0.500 0.522 0.511
cmt-edas 0.385 0.769 0.513
confof-edas 0.500 0.684 0.578
ekaw-sigkdd 0.538 0.636 0.583
confof-sigkdd 0.667 0.571 0.615
conference-sigkdd 0.588 0.667 0.625
conference-confof 0.550 0.733 0.629
confof-iasted 0.600 0.667 0.632
iasted-sigkdd 0.500 0.867 0.634
ekaw-iasted 0.533 0.800 0.640
edas-sigkdd 0.611 0.733 0.667
confof-ekaw 0.824 0.700 0.757
cmt-sigkdd 0.647 0.917 0.759
Average 0.506 0.653 0.555
Table 3. Results for conference dataset
3 General comments
3.1 Comments on the results
This is the first time that Eff2Match is participating in the OAEI competition and it has
shown good results compared to the results of other systems in the 2009 competition.
In particular, for the anatomy track, its F-Measure of 0.857 tops the best F-Measure
achieved in the OAEI 2009 anatomy track. What is more remarkable is that this was
achieved with a runtime of only around 2.5 minutes, thereby living up to its name of
being an effective and efficient matcher. In addition, its average F-Measure of 0.555 for
the conference track ranks second when compared with systems participating in OAEI
2009.
3.2 Discussions on ways to improve Eff2Match
The current implementation of Eff2Match only matches concepts and properties with
equivalence relations. The techniques used are mainly terminological and structural.
Our first proposed improvement to Eff2Match is to extend its functionalities to include
the discovery of non-equivalence correspondences. Other than the subsumption and
disjointedness correspondences defined in OWL, Eff2Match will also discover other
pre-defined relations that are common within the ontologies to be aligned. For example,
in the biomedical domain, many ontologies in the OBO foundry [1] contains the part-of relationship but they only connect concepts within the same ontology. We intend to
discover relationships like these in a future version of Eff2Match.
In addition, the current version of Eff2Match requires a few parameters to be set
by the user. The tuning of parameters is a manual process which can be tedious and
156
ineffective. Our other proposed improvement to Eff2Match is to enable it to tune the
parameters automatically, like what has been done in [2].
3.3 Comments on the OAEI 2010 procedure, test cases and measures
In this year’s OAEI competition, evaluation was done on the SEALS platform [4] for the
benchmark, anatomy and conference tracks. Matchers participating in this track have to
be implemented as web services so that they can be evaluated on the SEALS platform.
We feel that this service is very useful for us, particularly for the anatomy track. Unlike
the benchmark and conference tracks where the reference alignments are made known
to us, evaluation for the anatomy track is done using a blind test. Therefore, it is not
possible for us to observe how changes we make to the algorithm affect the results on
the anatomy dataset and it is difficult to make improvements. The SEALS platform has
alleviated this problem and made evaluation for the anatomy track possible.
Another useful feature of the SEALS evaluation mechanism is that it shows the
correct, incorrect and missing correspondences to the participants, allowing them to
gain a greater insight of the strengths and weaknesses of their systems. Though this is
an extremely useful feature, we feel that the correspondences for the anatomy dataset
should not be shown if it were to remain a blind test. The reason is that by joining the
set of correct correspondences and the set of missing correspondences, one can easily
get hold of the complete reference alignment for the anatomy dataset.
4 Conclusion
We have presented an ontology matcher named Eff2Match that can align ontologies
efficiently and effectively. Experiments were performed on different pairs of ontologies
from three different tracks in OAEI 2010 and results show that Eff2Match is able to
match real-world ontologies accurately. In addition, it scales well to large ontologies
and can be used in applications where the ontology matching process has to be fast.
References
1. The Open Biomedical Ontologies. Available at: http://obofoundry.org/about.shtml.
2. Mayssam Sayyadian, Yoonkyong Lee, AnHai Doan, and Arnon Rosenthal. Tuning schema
matching software using synthetic scenarios. In Proceedings of 31st International Conferenceon Very Large Data Bases (VLDB’05), pages 994–1005, Trondheim, Norway, 2005.
3. Nadine Schuurman and Agnieszka Leszczynski. Ontologies for bioinformatics. BioinformBiol Insights, 2:187–200, 2008.
4. SEALS-Semantic Evaluation At Large Scale. Available at:
http://seals.inrialpes.fr/platform.
5. Semantic Web for Earth and Environmental Terminology (SWEET). Available at:
http://sweet.jpl.nasa.gov/ontology/.
157
ObjectCoref & Falcon-AO: Results for OAEI 2010
Wei Hu, Jianfeng Chen, Gong Cheng, and Yuzhong Qu
1 Department of Computer Science and Technology, Nanjing University, China2 State Key Laboratory for Novel Software Technology, Nanjing University, China
{whu, yzqu}@nju.edu.cn
Abstract. In this report, we mainly present an overview of ObjectCoref, which
follows a self-training framework to resolve object coreference on the Semantic
Web. Besides, we show preliminary results of Falcon-AO (2010) for this year’s
OAEI campaign, including the benchmark and conference tracks.
1 Presentation of the system
1.1 State, purpose, general statement
The Semantic Web is an ongoing effort by the W3C Semantic Web Activity to actualize
data integration and sharing across different applications and organizations. To date,
a number of prominent ontologies have emerged to publish data for specific domains,
such as the Friend of a Friend (FOAF). These specifications recommend common iden-
tifiers for classes and properties in the form of URIs [1] that are widely and consistently
used across data sources.
On the instance level, however, it is far from achieving agreement among sources
on the use of common URIs to identify specific objects on the Semantic Web. In fact,
due to the decentralized and dynamic nature of the Semantic Web, it frequently happens
that different URIs from various sources, more likely originating from different RDF
documents, are used to identify the same real-world object, i.e., refer to an identical
thing (as known as URI aliases [5]). Examples exist in the domains of people, academic
publications, encyclopedic or geographical resources.
Object coreference resolution, also called consolidation or identification [2], is a
process for identifying multiple URIs of the same real-world object, that is, determin-
ing URI aliases (called coreferent URIs in this report) that denote a unique object. At
present, object coreference resolution is recognized to be useful for data-centric appli-
cations, e.g. heterogeneous data integration or mining systems, semantic search, query
and browsing engines.
We introduce a new approach, ObjectCoref, for bootstrapping object coreference
resolution on the Semantic Web. The architecture of the proposed approach follows a
common self-training framework (see Fig. 1). Self-training [6] is a major kind of semi-
supervised learning, which assumes that there are abundant unlabeled examples in the
real world, but the number of labeled training examples is limited. We believe that self-
training is an appropriate way for resolving object coreference on the Semantic Web.
Falcon-AO [4] is an automatic ontology matching system with acceptable to good
performance and a number of remarkable features. It is written in Java, and is open
158
source. ObjectCoref and Falcon-AO together help better enable interoperability be-
tween applications that use heterogeneous Semantic Web data.
Fig. 1. Self-training process
1.2 Specific techniques used
ObjectCoref builds an initial set of coreferent URIs mandated by the formal and explicit
semantics of owl:sameAs, owl:InverseFunctionalProperty, owl:FunctionalProperty,
owl:cardinality and owl:maxCardinality.
The semantics of owl:sameAs dictates that all the URIs linked with this property
have the same identity; if a property is declared to be inverse functional (IFP), then the
object of each property statement uniquely determines the subject (some individual);
a functional property (FP) is a property that can have only one unique value for each
object; while cardinality (or max-cardinality) allows the specification of exactly (or at
most) the number of elements in a relation, in the context of a particular class descrip-
tion, and when the number equals 1, it is somehow similar to the FP, but only applied
to this particular class.
Next, ObjectCoref learns the discriminability of pairs of properties based on the
coreferent URIs, in order to find more coreferent URIs for extending the training set.
The discriminability reflects how well each pair of properties can be used to determine
whether two URIs are coreferent or not. As an extreme example, IFPs (e.g. foaf:mbox)
have a very good discriminability.
In RDF graphs, each URI is involved in a number of RDF triples whose subject
is the URI, and the predicates and objects in these RDF triples form some <property,
value> pairs, which can be considered as features for describing such URI. ObjectCoref
compares the values between the <property, value> pairs from coreferent URIs, and
finds which two properties have similar values and how frequent. The significance is
the percentage of the number of coreferent URIs that can found by the discriminant
properties in all the coreferent URIs in the training set. If the significance is greater
than a given threshold, such the property pair is chosen for further resolution. Please
note that for different domains, same property pairs may have different discriminability.
159
For example, a pair of rdfs:labels is discriminant for the biomedical domain but not for
people.
If new coreferent URIs are found, ObjectCoref selects highly accurate ones and adds
them into the training set. The whole process iterates several times and terminates when
the property discriminability is not significant enough or cannot find more discriminant
property pairs.
1.3 Adaptations made for the evaluation
For ObjectCoref, there is no explicit equivalence semantics in the DI and PR tracks.
In order to establish the initial training set of coreferent URIs, we randomly extract 20
mappings from the reference alignment for each test case. All the mappings generated
by ObjectCoref are based on the same parameters.
For Falcon-AO, we do not make any specific adaptation in the OAEI 2010 cam-
paign. All the mappings for the benchmark and conference tracks outputted by Falcon-
AO are uniformly based on the same parameters.
1.4 Link to the system and parameters file
We implement an online service for ObjectCoref, and run it over a large-scale dataset
collected by the Falcons [3] search engine up to Sept. 2008. The dataset consists of
nearly 600 million RDF triples describing over 76 million URIs. It is still under devel-
opment. Please visit: http://ws.nju.edu.cn/objectcoref.
Besides, we follow the SEALS platform to publish Falcon-AO (2010) as a service.
Please access it from http://219.219.116.154:8083/falconWS?wsdl. The
offline version can be downloaded from our website: http://ws.nju.edu.cn/falcon-ao.
1.5 Link to the set of provided alignments (in align format)
The alignments for this year’s OAEI campaign should be available at the official web-
site: http://oaei.ontologymatching.org/2010/.
2 Results
In this section, we will present the results of ObjectCoref and Falcon-AO (2010) on the
tracks provided by the OAEI 2010 campaign.
2.1 DI
In this track, we use ObjectCoref to resolve object coreference between three pairs
of datasets, namely diseasome vs. sider, dailymed vs. sider and drugbank vs. sider. Ta-
ble 1 shows the discriminant property pairs that ObjectCoref learns by self-training. For
example, diseasome:name and sider:siderEffectName are a pair of discriminant proper-
ties, and if some URI in the diseasome dataset has a value w.r.t. diseasome:name that
160
is similar to some URI in the sider dataset w.r.t. sider:siderEffectName, these two URIs
can be considered as coreferent. In this track, the training process converges at two
iterations, respectively.
Table 1. Property discriminability on the DI track
Property in dataset1 Property in dataset2
diseasome vs. sider
rdfs:label sider:sideEffectName
diseasome:name sider:siderEffectName
rdfs:label rdfs:label
diseasome:name rdfs:label
dailymed vs. sider
dailymed:genericMedicine sider:drugName
dailymed:name sider:drugName
dailymed:genericMedicine rdfs:label
dailymed:name rdfs:label
drugbank vs. sider
drugbank:genericName sider:drugName
rdfs:label sider:drugName
drugbank:genericName rdfs:label
rdfs:label rdfs:label
drugbank:synonym sider:drugName
drugbank:synonym rdfs:label
drugbank:pubchemCompoundId sider:siderDrugId
drugbank:brandName sider:drugName
With these discriminant property pairs, ObjectCoref finds a number of coreferent
URIs for each pair of datasets. As shown in Table 2, the precision and recall is moderate.
Without considering the type of each object, the precision is not very good, so further
inference-based debugging on coreferent URIs is needed for future work.
Table 2. Performance of ObjectCoref on the DI track
Found Existing Precision Recall F-measure
diseasome vs. sider 190 238 0.837 0.668 0.743
dailymed vs. sider 2903 1592 0.548 0.999 0.708
drugbank vs. sider 933 283 0.302 0.996 0.464
2.2 PR
In this track, ObjectCoref uses the same self-training process to recognize coreferent
URIs for each pair of datasets, two of which are related to persons and the other is
about restaurants. The discriminant property pairs are listed in Table 3. Based on these
discriminant properties, ObjectCoref finds a set of coreferent URIs, where the precision
and recall are pretty good (see Table 4). In particular, the good recall reflects that our
161
learning approach identifies the key properties for resolving object coreference in this
track. But we also notice that some combination of properties may be also helpful. For
example, first name + last name can be used for identifying same people.
Table 3. Property discriminability on the PR track
Property in dataset1 Property in dataset2
person1
person11:has address person12:has address
person11:phone number person12:phone number
person11:soc sec id person12:soc sec id
person2
person21:has address person22:has address
person21:phone number person22:phone number
person21:soc sec id person22:soc sec id
restaurantsrestaurant1:has address restaurant2:has address
restaurant1:name restaurant2:name
Table 4. Performance of ObjectCoref on the PR track
Found Existing Precision Recall F-measure
person1 499 500 1.000 0.998 0.999
person2 362 400 1.000 0.900 0.947
restaurants 91 112 0.989 0.804 0.887
2.3 Benchmark & conference
We use Falcon-AO (2010) to participate in the benchmark and conference tracks. The
average precision and recall are depicted in Table 5. As compared to OAEI 2007, the
benchmark track adds some new cases. Falcon-AO failed in several cases due to the
Jena parsing errors. For the detailed results, please see Appendix.
Table 5. Performance of Falcon-AO (2010) on the benchmark and conference tracks
Precision Recall
Benchmark 0.76 0.64
Conference 0.60 0.60
3 General comments
In this section, we will firstly discuss several possible ways to improve ObjectCoref,
and then give comments on the OAEI 2010 test cases.
162
3.1 Discussions on the way to improve the proposed system
The preliminary results of ObjectCoref demonstrate that using property discriminability
is feasible to find coreferent URIs on the Semantic Web. However, we also see several
shortcomings of the proposed approach, which will be considered in the next version.
1. How to divide objects into different domains? For the tasks in this year’s OAEI,
we may not see the importance of recognizing domains, but on the whole Seman-
tic Web, different domains may have different discriminant properties, and a single
property pair may have different discriminability in different domains. So, a uni-
form measurement is ineffective.
2. How to avoid error accumulation? In self-training, an important issue is to prevent
error accumulation, since a wrong labeled example would lead to misclassification
in further propagation. In our evaluation, because the training process converges in
a few iterations, so this situation is not so significant. But in real world, it is imper-
ative to consider that.
3. How to find discriminant property combinations? A single property may be not
good enough for resolving object coreference, while the combination of several
properties would be more discriminant. However, we need to avoid overfitting. So,
we plan to mine frequent patterns in the RDF data for describing objects and refine
these frequent patterns to form property combinations.
3.2 Comments on the OAEI 2010 test cases
The proposed matching tasks cover a large portion of real world domains, and the dis-
crepancies between them are significant. Doing experiments on these tasks are helpful
to improve algorithms and systems. In order to enhance applicability, we list some prob-
lems in our experiment procedure, which might aid organizers to improve in the future.
1. In the DI track, the organizers provide 4 downloadable datasets for the biomedi-
cal domain, however, the interlinking track also involves a number of others, e.g.,
linkedct, lifescience, bio2rdf. The datasets are not only very large, but also diffi-
cult to find the latest versions, most of which are even not allowed to download.
Furthermore, using SPARQL endpoints in the experiment is very time-consuming,
especially for such a large scale. So, we would expect that all the datasets can be
(perhaps temporarily) offline in the next year.
2. Falcon-AO (2010) uses Jena 2.6.3 as the RDF parser. In the benchmark track, some
ontologies may have problems and cause the Jena exception “Unqualified typed
nodes are not allowed. Type treated as a relative URI”. So, we would expect the
organizers to fix this in the next year.
163
4 Conclusion
Object coreference resolution is an important way for establishing interoperability among
(Semantic) Web applications that use heterogenous data. We implement an online sys-
tem for resolving object coreference called ObjectCoref, which follows a self-training
framework focusing on learning property discriminability. From the experiments in this
year’s DI and PR tracks, we find some positive and negative experience for improving
our system. In the near future, we look forward to making a stable progress towards
building a comprehensive object coreference resolution system for the Semantic Web.
Acknowledgements
This work is in part supported by the NSFC under Grant 61003018 and 60773106. We
would like to thank Ming Li for his valuable comments on self-training.
References
1. Berners-Lee, T., Fielding, R., Masinter, L.: Uniform Resource Identifier (URI): Generic Syn-
tax. RFC 2396 http://www.ietf.org/rfc/rfc3986.txt
2. Bleiholder, J., Naumann, F.: Data Fusion. ACM Computing Surveys 41(1), 1–41 (2008)
3. Cheng, G., Qu, Y.: Searching Linked Objects with Falcons: Approach, Implementation and
Evaluation. International Journal on Semantic Web and Information Systems 5(3): 49–70
(2009)
4. Hu, W., Qu, Y.: Falcon-AO: A Practical Ontology Matching System. Journal of Web Seman-
tics 6(3), 237–239 (2008)
5. Jacobs, I., Walsh, N.: Architecture of the World Wide Web, Volume One. W3C Recommen-
dation 15 December 2004. http://www.w3.org/TR/webarch/6. Zhou, Z., Li, M.: Semi-supervised learning by disagreement. Knowledge and Information
Systems 24(3), 415–439 (2009)
Appendix: Complete results
In this appendix, we will show the complete results of Falcon-AO (2010) on the bench-mark and conference tracks. Tests were carried out on two Intel Xeon Quad 2.40GHz
CPUs, 8GB memory with Redhat Linux Enterprize Server 5.4 (x64), Java 6 compiler
and MySQL 5.0.
Matrix of Results
In the following tables, the results are shown by precision (Prec.) and recall (Rec.).
164
Bench Description Prec. Rec.
101 Reference 1.00 1.00102 Irrelevant NaN NaN103 Lang. generalization 1.00 1.00104 Lang. Restriction 1.00 1.00201 No names 0.97 0.97201-2 0.98 0.98201-4 1.00 1.00201-6 0.92 0.92201-8 0.98 0.98202 No names, comments Jena error202-2 0.70 0.70202-4 0.70 0.70202-6 0.72 0.72202-8 0.72 0.72203 Misspelling 1.00 1.00204 Naming conventions 0.96 0.96205 Synonyms 0.97 0.97206 Translation 0.94 0.94207 0.96 0.96208 0.98 0.98209 0.65 0.65210 0.68 0.68221 No specialization 1.00 1.00222 Flattened hierarchy 1.00 1.00223 Expanded hierarchy 1.00 1.00224 No instances 1.00 0.99225 No restrictions 1.00 1.00228 No properties 1.00 1.00230 Flattened classes 0.94 1.00231 Expanded classes 1.00 1.00232 1.00 0.99233 1.00 1.00236 1.00 1.00237 1.00 0.99238 1.00 0.99239 1.00 1.00240 1.00 1.00241 1.00 1.00246 1.00 1.00247 1.00 1.00248 Jena error248-2 0.69 0.68248-4 0.71 0.69248-6 0.73 0.71248-8 0.69 0.68249 Jena error249-2 0.70 0.70249-4 0.71 0.71249-6 0.73 0.73249-8 0.73 0.73250 Jena error250-2 1.00 0.79250-4 1.00 0.61250-6 0.93 0.42250-8 0.88 0.21
Bench Description Prec. Rec.
251 Jena error251-2 0.99 0.78252-4 0.53 0.53252-6 0.55 0.55252-8 0.55 0.55253 Jena error253-2 0.71 0.69253-4 0.69 0.68253-6 0.67 0.66253-8 0.69 0.68254 Jena error254-2 1.00 0.79254-4 1.00 0.61254-6 0.93 0.42254-8 0.88 0.21257 Jena error257-2 1.00 0.79257-4 1.00 0.61257-6 0.93 0.42257-8 0.88 0.21258 Jena error258-2 0.99 0.78258-4 1.00 0.59258-6 0.97 0.40258-8 0.95 0.22259 Jena error259-2 0.55 0.55259-4 0.51 0.51259-6 0.55 0.55259-8 0.54 0.54260 Jena error260-2 0.96 0.79260-4 1.00 0.62260-6 0.92 0.41260-8 0.88 0.24261 Jena error261-2 1.00 0.79261-4 1.00 0.79261-6 1.00 0.79261-8 1.00 0.79262 Jena error262-2 1.00 0.79262-4 1.00 0.61262-6 0.93 0.42262-8 0.88 0.21265 Jena error266 Jena error301 BibTeX/MIT 0.91 0.68302 BibTeX/UMBC 0.87 0.56303 BibTeX/Karlsruhe 0.73 0.73304 BibTeX/INRIA 0.95 0.92
Conference Prec. Rec.
cmt-conference 0.50 0.56cmt-confof 0.55 0.38cmt-edas 0.69 0.69cmt-ekaw 0.55 0.55cmt-iasted 0.50 1.00cmt-sigkdd 0.77 0.83conference-confof 0.53 0.60conference-edas 0.48 0.59conference-ekaw 0.50 0.48conference-iasted 0.63 0.36conference-sigkdd 0.71 0.67confof-edas 0.45 0.53confof-ekaw 0.62 0.65confof-iasted 0.36 0.44confof-sigkdd 0.80 0.57edas-ekaw 0.65 0.57edas-iasted 0.64 0.37edas-sigkdd 0.88 0.47ekaw-iasted 0.54 0.70ekaw-sigkdd 0.78 0.64iasted-sigkdd 0.59 0.87
165
An integrated matching system: GeRoMeSuite and SMBResults for OAEI 2010
Christoph Quix1, Avigdor Gal2, Tomer Sagi2, David Kensche1
1RWTH Aachen University, Germanyhttp://www.dbis.rwth-aachen.de
2Technion – Israel Institute of Technologyhttp://www.technion.ac.il/
Abstract. We present the results of an integrated matching system which is theresult of a cooperation project between the Israel Institute of Technology (Tech-nion) and the RWTH Aachen University in Germany. We have integrated theGeRoMeSuite system (from RWTH Aachen) and SMB (from Technion). Bothtools aim at matching schemas; while GeRoMeSuite offers a variety of matchers,SMB provides the information on how to combine matchers and how to enhancematch results. Thus, an integration of the tools is beneficial for both systems.
1 Presentation of the system
1.1 GeRoMeSuite
As a framework for model management, GeRoMeSuite [3] provides an environment tosimplify the implementation of model management operators. GeRoMeSuite is based onthe generic role based metamodel GeRoMe [2], which represents models from differentmodeling languages (such as XML Schema, OWL, SQL) in a generic way. Thereby,the management of models in a polymorphic fashion is enabled, i.e., the same operatorimplementations are used regardless of the original modeling language of the schemas.In addition to providing a framework for model management, GeRoMeSuite implementsseveral fundamental operators such as Match [6], Merge [5], and Compose [4].
The matching component of GeRoMeSuite has been described in more detail in[6], where we present and discuss in particular the results for heterogeneous matchingtasks (e.g., matching XML Schema and OWL ontologies). An overview of the completeGeRoMeSuite system is given in [3].
1.2 SMB
The Schema Matching Boosting (SMB) Service is a toolkit for enhancing the perfor-mance of schema matchers. SMB operates in 3 modes: Enhance, Learn, and Recom-mend. In the enhance mode, SMB recieves a raw correspondence matrix (with similar-ity values for attribute correspondence in the range of [0,1]) and performs an analysisof the results per row and column. Subsequently, SMB uses contrasting and weaken-ing algorithms to boost results of “promising” rows and columns and weaken results
166
of “non-promising” rows and columns respectively. Contrasting is perfromed using amodified version of the Weber contrast function. Weakening is inversly proportional tothe row and column average.
The learn mode is used to perform off-line training of SMB on the perfromancebehavior of matchers w.r.t. various matching tasks which are classified to classes ac-cording to their a-priory features such as schema size. Training is performed using theSMB algorithm, as introduced in [1]. The recommend classifies in run-time a givenmatching task, providing the reccomended ensemble weights for the matching systemsvarious components. The Learn and recommend modes are a re-implementation of thesystem presented in [1] in which run-time complexity has been reduced from O(n!) toO(n2) and generic interfaces have been provided to allow any matching system to useSMB by command-line invocation.
1.3 State, purpose, general statement
GeRoMeSuite is a generic system which can match ontologies as well as schemas inother modeling languages such as XML Schema or SQL. Therefore, it is well suitedfor matching tasks across heterogeneous modeling languages, such as matching XMLSchema with OWL. We discussed in [6] that the use of a generic metamodel, whichrepresents the semantics of the models to be matched in detail, is more advantageousfor such heterogeneous matching tasks than a simple graph representation.
SMB is also a modeling language independent ‘meta’ matching system which mainlyworks on the similarity matrices produced by GeRoMeSuite. It improves the clarity ofthe similarity values by improving ‘good’ values and descreasing ‘bad’ values. Thisshould increase the precision of the match result.
1.4 Specific techniques used
Besides the integration of GeRoMeSuite and SMB, we focused this year on adding val-idation methods to the system to improve the precision of the match result. A compo-nent for adding disjointness relationships in an ontology has been added to the match-ing framework. The component uses machine learning techniques to identify disjointconcepts with one ontology. The disjointness relationships can then be used in the vali-dation of schema matches using logical reasoning.
Furthermore, we developed a component which can use a background ontologyto find additional matches in the ontology. The system is able to find an appropriatebackground ontology on the web automatically, using Google and Swoogle. Due to theset up of the OAEI campaign, we did not use this component for OAEI.
1.5 Adaptations made for the evaluation
We evaluated several match configurations which is easily possible due to the adaptableand extensible matching framework of GeRoMeSuite. As only one configuration canbe used for all matching tasks, we had to find a good compromise between performancein terms of precision and recall, time performance for larger ontologies (e.g., anatomy),
167
and selection of appropriate matchers which work well on all tracks. For example, wealso tested configurations which had an f-measure that was about 5% higher than theconfiguration which we eventually used, but these configurations did not work well onall tracks. The identification of good match configurations is a topic for future research.
Fig. 1 indicates the strategy which we used for the matching tasks in the benchmarktrack. All aggregation and filter steps use variable weights and thresholds, which arebased on the statistical values of the input similarities.
The role matcher is a special matcher which compares the roles of model elements inour generic role-based metamodel. In principle, this results in matching only elementsof the same type, e.g., classes with classes only and properties with properties only.
Stringatcher
ChildrenMatcher
ParentMatcher
Aggregation
RoleMatcher
AggregationFinalFilterstance
atcher
Aggregation
MatchValidation
SMB Enhancer
SMB Enhancer
StringMatcher
ChildrenMatcher
ParentMatcher
Aggregation
RoleMatcher
AggregationFinalFilterInstance
Matcher
Aggregation
MatchValidation
SMB Enhancer
SMB Enhancer
Fig. 1. Matching Strategy for OAEI 2010
On a technical level, we implemented a command line interface for the match-ing component, as the matching component is normally used from within the GUI ofGeRoMeSuite. The command line interface can work in a batch mode in which severalmatching tasks and configurations can be processed and compared. The existence ofthis tool enabled also an easy integration with the OAEI web service interface.
1.6 Link to the system and parameters file
More information about the system can be found on the homepage of GeRoMeSuite:http://www.dbis.rwth-aachen.de/gerome/oaei2010/
The page provides also links to the configuration files used for the evaluation.
1.7 Link to the set of provided alignments (in align format)
The results for the OAEI campaign 2010 are available at http://www.dbis.rwth-aachen.de/gerome/oaei2010/
2 Results
2.1 Benchmark
The following table shows the average results for precision and recall in the benchmarktrack.
168
Task Precision Recall1xx 1,00 1,002xx (xx<48) 0,96 0,882xx (xx>47) 0,89 0,513xx 0,90 0,42A first check, whether a match configuration is suitable at all are the 1xx ontologies.
A configuration should produce the perfect result for these tracks, which is the case forthe configuration, we have finally chosen.
For the simpler tasks in the 2xx data set (201-247), our system was able to achievea very good result with an f-measure of more than 0.9.
For some of the really difficult tasks (248-266), our system was not able to find anycorrespondence as there is hardly any information that can be used (e.g., task 265 withno labels, no comments, no hierarchy, etc.).
The results for the tasks 3xx was in general good (f-measure of about 0.6 for 301,302, and 304). However, ontology 303 is difficult for our generic system as the names-paces are not defined in a standard way. Therefore, we could only find a few correspon-dences.
2.2 Conference
The ontologies in the conference track are rather small and the matching tasks are moredifficult as the ontologies have been designed by humans using different terminologiesand having different goals in mind. As this is a more realistic case than the benchmarktrack, we have chosen a configuration which produces good results for the conferencetrack. Using validation rules to check the logical consistency of the identified correspon-dences and a final filter step which generates only 1:1 correspondences was beneficialfor the quality of the result.
We achieve an average f-measure of about 45%, which is rather low compared toother systems. Low precision and the lack of coherence in the produced alignmentsindicates that we need to improve the validation methods for the computed alignments,such that more false positives (i.e., incoherent correspondences) are removed.
2.3 Anatomy
We participated in this task in the sub-tracks 1 to 3. Because of some special featuresin the input ontologies (e.g., synonyms are related to classes with special object projectproperties), our system is not able to detect most correspondences in the standard con-figuration. We achieved a high precision but low recall in task 1. Therefore, we used theresult of task 1 also as a result for task 2. In task 3, we achieved a high recall by usinga configuration with very low thresholds and without any validation methods.
2.4 Directory
We participate only in the single task modality of the directory track. The size of theinput ontologies is similar to the anatomy track, so the same problems of scalabilityhave to be faced here. We submitted an alignment with about 700 correspondences.Unfortunately, the reference alignment could not be produced for the single track.
169
3 Comments
We participate this time the third time in OAEI and see again some improvement ofour matcher compared to last year. Thus, a structured evaluation and comparison ofontology alignment and schema matching components as OAEI is very useful for thedevelopment of such technologies. We appreciate especially the automatic evaluationsystem, although we also had to put some additional effort to get the interface and ourweb service working.
We are currently working on a system to generate a matching benchmark whichcomes closer to the challenges of real ontologies. We would be happy if we could con-tribute the results to OAEI 2011. Furthermore, an oriented track as in OAEI 2009 wouldbe useful to evaluate semantic matching techniques.
4 Conclusion
As our tool is neither specialized on ontologies nor limited to the matching task, we didnot expect to deliver the best results. However, we are very satisfied with the overallresults, as we can compete with the special purpose ontology alignment tools.
We will continue to work on the improvement of our matching system and on theintegration of GeRoMeSuite and SMB. We will especially focus on the problem ofidentifying good match configurations automatically. We hope to participate again withan improved system in the OAEI campaign next year.Acknowledgements: This work is supported by the DFG Research Cluster on UltraHigh-Speed Mobile Information and Communication (UMIC, http://www.umic.rwth-aachen.de) and by the Umbrella Cooperation Programme (http://www.umbrella-coop.org/).
References
1. A. Gal, T. Sagi. Tuning the Ensemble Selection Process of Schema Matchers. InformationSystems, 35(8):845–859, 2010.
2. D. Kensche, C. Quix, M. A. Chatti, M. Jarke. GeRoMe: A Generic Role Based Metamodelfor Model Management. Journal on Data Semantics, VIII:82–117, 2007.
3. D. Kensche, C. Quix, X. Li, Y. Li. GeRoMeSuite: A System for Holistic Generic ModelManagement. C. Koch, J. Gehrke, M. N. Garofalakis, D. Srivastava, K. Aberer, A. Deshpande,D. Florescu, C. Y. Chan, V. Ganti, C.-C. Kanne, W. Klas, E. J. Neuhold (eds.), Proceedings33rd Intl. Conf. on Very Large Data Bases (VLDB), pp. 1322–1325. Vienna, Austria, 2007.
4. D. Kensche, C. Quix, Y. Li, M. Jarke. Generic Schema Mappings. Proc. 26th Intl. Conf. onConceptual Modeling (ER’07), pp. 132–148. 2007.
5. C. Quix, D. Kensche, X. Li. Generic Schema Merging. J. Krogstie, A. Opdahl, G. Sin-dre (eds.), Proc. 19th Intl. Conf. on Advanced Information Systems Engineering (CAiSE’07),LNCS, vol. 4495, pp. 127–141. Springer-Verlag, 2007.
6. C. Quix, D. Kensche, X. Li. Matching of Ontologies with XML Schemas using aGeneric Metamodel. Proc. Intl. Conf. Ontologies, DataBases, and Applications of Seman-tics (ODBASE), pp. 1081–1098. 2007.
170
LN2R – a knowledge based reference reconciliationsystem: OAEI 2010 Results
Fatiha Saıs1, Nobal Niraula2, Nathalie Pernelle1, Marie-Christine Rousset3
1 LRI (Paris-Sud 11 University & CNRS) / INRIA Saclay Parc-Club Orsay Univ.
Bat. G ,4, rue Jacques Monod F-91893 Orsay Cedex, France
{fatiha.sais, nathalie.pernelle}@lri.fr2 University of Memphis, TN, USA
[email protected] LIG - Laboratoire dInformatique de Grenoble
BP 72, 38402 St MARTIN DHERES, France
Abstract. This paper presents the first participation of LN2R system in
IM@OAEI2010, the Instance Matching track of Ontology Alignment Evaluation
Initiative 2010 Campaign. In particular, we participated in OWL data track by
performing LN2R system on Person-Restaurant data set. We obtained very good
results on person data sets and reasonable results on restaurant data set.
1 Presentation of the system
To design a semantic information integration system, we are faced to two reconciliation
problems. First, the schema (or ontology) reconciliation which consists in finding map-
pings between elements (concepts or relations) of two schemas or two ontologies (see
[1, 2] for surveys). The second problem concerns data reconciliation (named reference
reconciliation) which consists in comparing data descriptions and deciding whether dif-
ferent descriptions refer to the same real world entity (e.g. the same person, the same
article, the same gene). The problem of reference reconciliation is very critical, since it
impacts data quality and data consistency [3].
In LN2R system, we address only the problem of reference reconciliation. There
are several kinds of reference reconciliation approaches: knowledge-based, similarity-
based, probabilistic, supervised, etc.[4]. In this paper we focus our study on reference
reconciliation approaches that are informed and global. Informed approaches are those
which exploit knowledge that is declared in the ontology to reconcile data. Reference
reconciliation approaches are said global when they exploits the dependencies possibly
existing between reference reconciliations [5, 6]. Such approaches use attribute values
describing the data but also references that are related to the considered data [5, 6]. For
example, the reconciliation between two scientists can entail the reconciliation between
their two affiliated universities. Such dependencies result from the semantics of the
domain of interest.
171
1.1 State, purpose, general statement
The reference reconciliation system (LN2R) that we have tested in IM@OAEI2010
campaign is knowledge-based, unsupervised and based on two methods, a logical one
called L2R and a numerical one called N2R. The Logical method for Reference Rec-
onciliation (L2R) is based on the translation in first order logic (Horn rules) of some
of the schema semantics. In order to complement the partial results of L2R, we have
designed a Numerical method for Reference Reconciliation (N2R). It exploits the L2R
results and allows computing similarity scores for each pair of references.
Reference reconciliation problem. Let S1 and S2 be two data sources which con-
form to the same OWL ontology. Let I1 and I2 be the two reference sets that cor-
respond respectively to the data of S1 and S2. The problem consists in deciding
whether references are reconciled or not reconciled. Let Reconcile be a binary predi-
cate. Reconcile(X,Y ) means that the two references denoted by X and Y refer to the
same world entity. The reference reconciliation problem considered in L2R consists in
extracting from the set I1× I2 of reference pairs two subsets REC and NREC such
that: REC = {(i, i), Reconcile(i, i)} and NREC = {(i, i),¬Reconcile(i, i)}The reference reconciliation problem considered in N2R consists in, given a simi-
larity function Simr : I1× I2 → 0..1, and a threshold Trec (a real value in 0..1 given
by an expert, fixed experimentally or learned on a labeled data sample), computing the
following set:
RECN2R = {(i, i′) ∈ (I1 × I2)\(REC ∪NREC), s.t.Simr(i, i′) > Trec}
1.2 Specific techniques used
In the following, we will present some details on the knowledge-based reference recon-
ciliation system (LN2R). First, we will show through an example the ontology and the
kind of knowledge that we use. Second, we give a brief presentation of the two methods
L2R and N2R of reference reconciliation.
The ontology and its constraints The considered OWL ontology consists of a set
of classes (unary relations) organized in a taxonomy and a set of typed properties (bi-
nary relations). These properties can also be organized in a taxonomy of properties.
Two kinds of properties can be distinguished in OWL: the so-called relations (in OWL
abstractProperty), the domain and the range of which are classes and the so-called at-
tributes (in OWL objectProperty), the domain of which is a class and the range of which
is a set of basic values (e.g. Integer, Date, Literal).
We allow the declaration of constraints expressed in OWL-DL or in SWRL in order
to enrich the domain ontology. The constraints that we consider are of the following
types:
– Constraints of disjunction between classes: DISJOINT(C,D) is used to declare that
the two classes C and D are disjoint.– Constraints of functionality of properties: PF(P) is used to declare that the property
P (relation or attribute) is a functional property.
172
(a) (b)
Source S1:MuseumName(S1 m1,”Le Louvre”); Con-tains(S1 m1,S1 p1); Located(S1 m1,S1 c1);CityName(S1 c1,”Paris”); PaintingName(S1 p1,”La Joconde”);
Source S2:MuseumName(S2 m1,”muse du Louvre”);Located(S2 m1,S2 c1); Contains(S2 m1,S2 p1);Contains(S2 m1,S2 p2);CityName(S2 c1,”Villede paris”); PaintingName(S2 p1, ”Abricotiers enfleurs”); PaintingName(S2 p2,”Joconde”);
Fig. 1. (a) an extract of cultural place ontology, (b) an extract of RDF data
– Constraints of inverse functionality of properties: PFI(P) is used to declare that the
property P (relation or attribute) is an inverse functional property. These constraints
can be generalized to a set {P1, . . . , Pn} of relations or attributes to state a com-
bined constraint of inverse functionality that we will denote PFI(P1, . . . , Pn). For
example, PFI(located, name) expresses that one address and one name cannot
be associated to several cultural places (i.e. both are needed to identify a cultural
place).
Data description and their constraints A piece of has a reference, which has the
form of a URI (e.g. http://www.louvre.fr,NS-S1/painting243), and a
description, which is a set of RDF facts involving its reference. An RDF fact can be:
either (i) a class-fact C(i), where C is a class and i is a reference, (ii) a relation-fact
R(i1, i2), where R is a relation and i1 and i2 are references, or (iii) an attribute-fact
A(i, v), where A is an attribute, i a reference and v a basic value (e.g. integer, string,
date).
The data description that we consider is composed of the RDF facts coming from
the data sources enriched by applying the OWL entailment rules. We consider that the
descriptions of data coming from different sources conform to the same OWL ontology
(possibly after schema reconciliation). Figure 1 (b), provides examples of data coming
from two RDF data sources S1 and S2, which conform to a same ontology describing
the cultural application previously mentioned.
L2R: a Logical method for Reference Reconciliation L2R [7] is based on
the inference of facts of reconciliation (Reconcile(i,j) ) and of non-reconciliation
(¬Reconcile(i′, j′)) from a set of facts and a set of rules which transpose the
semantics of the data sources and of the schema into logical dependencies between
reference reconciliations. Facts of synonymy (SynVals(v1,v2)) and of no synonymy
(¬SynV als(u1, u2)) between basic values (strings, dates) are also inferred. For in-
stance, the synonymy SynVals(“JoDS”, “Journal of Data Semantics”) may be inferred.
173
The L2R distinguishing features are that it is global and logic-based: every con-
straint declared on the data and on the OWL ontology is automatically translated into
first-order logic Horn rules (rules for short) that express dependencies between reconcil-
iations. For instance, the following rule R translates the knowledge that the two classes
Museum and City are disjoint, R :Museum(X) ∧ City(Y ) ⇒ ¬Reconcile(X,Y )
To deduce all the (non ) reconciliation and (non) synonymy facts from the knowl-
edge base, we use a logical reasoning based on the unit-resolution inference rule. The
advantage of such a logical approach is that if the data are error-free and if the declared
constraints are valid, then the reconciliations and non-reconciliations that are inferred
are correct, thus guaranteeing a 100 % precision of the results.
N2R: a Numerical method for Reference Reconciliation N2R [5] has two main dis-
tinguishing characteristics. First, it is fully unsupervised: it does not require any train-
ing phase from manually labeled data to set up coefficients or parameters. Second, it is
based on equations that model the influence between similarities. In the equations, each
variable represents the (unknown) similarity between two references while the simi-
larities between values of attributes are constants. These constants are obtained, either
(i) by exploting a dictionnary of synonyms (e.g. WordNet thesaurus, the dictionnary
of synonyms generated by L2R [7]); or (ii) by using standard similarity measures on
strings or on sets of strings. Furthermore, ontology and data knowledge (disjunctions
and UNA) is exploited by N2R in a filtering step to reduce the number of reference
pairs that are considered in the equation system. The functions modeling the influence
between similarities are a combination of maximum and average functions in order to
take into account the constraints of functionality and inverse functionality declared in
the OWL ontology in an appropriate way.
The equations modeling the dependencies between similarities. For each pair of
references, its similarity score is modeled by a variable xi and the way it depends on
other similarity scores is modeled by an equation: xi = fi(X), where i ∈ [1..n] and nis the number of reference pairs for which we apply N2R, and X = (x1, x2, . . . , xn).Each equation xi = fi(X) is of the form: fi(X) = max(fi−df (X), fi−ndf (X))
The function fi−df (X) is the maximum of the similarity scores of the value pairs
and the reference pairs of attributes and relations with which the i-th reference pair
is functionally dependent. The maximum function allows propagating the similarity
scores of the values and the references having a strong impact. The function fi−ndf (X)is defined by a weighted average of the similarity scores of the values pairs (and sets)
and the reference pairs (and sets) of attributes and relations with which the i-th reference
pair is not functionally dependent. See [5] for the detailed definition of fi−df (X) and
fi−ndf (X).
Iterative algorithm for reference pairs similarity computation. Solving this equation
system is done by an iterative method inspired from the Jacobi method [8], which is
fast converging on linear equation systems. To compute the similarity scores, we have
implemented an iterative resolution method. At each iteration, the method computes
the variable values by using those computed in the precedent iteration. Starting from an
initial vector X0 = (x01, x02, ..., x
0n), the value of the vector X at the k-th iteration is
174
obtained by the expression:Xk = F (Xk−1). At each iteration k we compute the value
of each xki : xki = fi(xk−11 , xk−1
2 , ...xk−1n ) until a fix-point with precision ε is reached.
The fix-point is reached when: ∀i, |xki − xk−1i | <= ε.
The similarity computation is illustrated by the equation system (see Table 1) ob-
tained from the data descriptions shown in the example 1.
– x1 = Simr(S1 m1, S2 m1) ; Simv(“Le louvre”, “Musee du louvre”) = 0.68– x2 = Simr(S1 p1, S2 p1) ; Simv(“La Joconde”, “Abricotiers en fleurs”) = 0.1– x3 = Simr(S1 p1, S2 p2) ; Simv(“La Joconde”, “Joconde”) = 0.9– x4 = Simr(S1 c1, S2 c1) ; Simv(“Paris”, “Ville de Paris”) = 0.42
The weights are computed in function of the number of common attributes and
common relations of the reference pairs. The weights used in the value computation of
the variables x1, x2, x3 and x4 are respectively: λ11 = 1/4, λ21 = 1/2, λ31 = 1/2 and
λ41 = 1/2. We assume that point-fix precision ε is equal to 0.005.
The equation system is the one given in the example 2. The different iterations of
the resulting similarity computation are provided in Table 1.
Iterations 0 1 2 3 4x1 = max(0.68, x2, x3,
14∗ x4) 0 0.68 0.9 0.9 0.9
x2 = max(0.1, 12∗ x1) 0 0.1 0.34 0.45 0.45
x3 = max(0.9, 12∗ x1) 0 0.9 0.9 0.9 0.9
x4 = max(0.42, x1) 0 0.42 0.68 0.9 0.9
Table 1. Example of iterative similarity computation
The solution of the equation system is X = (0.9, 0.45, 0.9, 0.9). This corresponds
to the similarity scores of the four reference pairs. The fix-point has been reached after
four iterations. If we fix the reconciliation threshold Trec at 0.80, then we obtain three
reconciliation decisions: two cities, two museums and two paintings.
1.3 Adaptations made for the evaluation
In order to perform the evaluation of LN2R system on the data sets provided by
IM@OAEI2010 evaluation campaign we were faced to do some choices and adapta-
tions. As LN2R system assume that the data sets conform to the same ontology, we
have performed the following steps:
1. manual alignment of the two ontologies (schemas),2. choose a federated ontology (one among the two considered ontologies) and3. transform the other ontology to the chosen one.
Thanks to the small size of the considered ontologies and to their structural and
semantic closeness, the above steps were performed easily. For example, for the restau-
rant data set, the difference between the two ontologies is that the category and cityare object properties in one ontology and data properties in the other.
In addition to this, we have transformed the LN2R output to comply with the OAEI
alignment format.
175
1.4 Link to the system and parameters file
The LN2R system (including the parameters file) can be downloaded at:
http://www.lri.fr/˜sais/IM-OAEI10/LN2RSystem.zip.
1.5 Link to the set of provided alignments (in align format)
The results that are obtained by LN2R in the instance matching track of OAEI 2010
campaign can be found at:
http://www.lri.fr/˜sais/IM-OAEI10/LN2RResults.zip.
2 Results
In this section we present our comments on the results obtained from the first partici-
pation of LN2R in the Instance matching track IM@OAEI 2010. We have tested LN2R
system on person and restaurant (PR) data sets.
To evaluate our system we have compared its results on the different data sets with
the provided gold-standard and we have computed the recall, the precision and the F-
measure.
2.1 Person1 and Person2 data sets
In this track, participants are asked to find all correct alignments between person’ in-
stances for the two data sets person1 and person2. Each data set contains the OWL
ontologies, the two RDF files to be reconciled and the reference alignments (gold-
standard) file.
Since, in LN2R, the reconciliation decisions are based on a reconciliation threshold,
we have performed several tests by varying the threshold value from 0.6 to 1. The best
results are obtained for a threshold of 0.75 for both data sets: (i) for person1 data set, we
have obtained the maximum F-measure of 100 % for a recall of 100 % and a precision
of 100% and (ii) for the person2 data set we have obtained a F-measure of 93% for a
recall of 88.25 % and a precision of 99.4 %.
Furthermore, as our method is global in the sense that the reconciliation decisions
between instances are propagated to other pairs of instances through the relations which
link them together, we have also inferred alignments between address instances. Thanks
to the functionnality of has − address property, we have obtained 500 alignments
between address instances for person1 data set and 355 alignments for person2 data
set for a threshold of 0.75.
In addition to reconciliations (positive alignments), we also infer non reconciliations
between instances, thanks to the reasoning on ontology knowledge, like disjunctions
and UNA. In person1 and person2 ontologies we have declared the disjunction be-
tween person and address classes which leads to the inference of a non reconciliation
between every person instance and every address instance. These non reconciliations
are very useful in a filtering step where unnecessary comparisons and similarity com-
putation are avoided.
176
2.2 Restaurant data set
In this track, participants are asked to find all correct alignments between restaurant’
instances of the restaurant1 data set. It contains the OWL ontologies, the two RDF
files to be reconciled and the reference alignments (gold-standard) file.
As we have done for the person data sets, we have also performed several tests of the
system by varying the threshold value from 0.6 to 1. Comparing to the gold standard,
the best results are obtained for a threshold of 0.85. We have obtained a F-measure of
75.3 % for a recall of 75 % and a precision of 75.67 %.
Similarly to the person data sets, we have also inferred a set of alignments between
address instances thanks to the propagation mechanism. In a filtering step, we have
also inferred a set of non reconciliation (negative alignments) between restaurant and
address instances.
By analyzing the results of the restaurant data set, we have noticed some mistakes
in the provided reference alignments: correct alignments are missed (see example 1)
and some given alignments are wrong (see example 2).
Example 1: the two instances (http://www.okkam.org/oaie/restaurant1-Restaurant16,
http://www.okkam.org/oaie/restaurant2-restaurant26) should be included in the gold-
standard because they refer to the same restaurant. There descriptions are as follows:
(1) [’name: patina’, ’category: californian’, ’phone number:213/467-1108’ , has-
address[street:’5955 melrose ave.’, is in city[name:los angeles’]]]
(2) [’name: patina’, has category[’name:californian’], ’phone number:213-467-1108’ ,
has-address[’city:los angeles’, street:’5955 melrose ave.’]]
Example 2: the two instances (http://www.okkam.org/oaie/restaurant1-Restaurant2,
http://www.okkam.org/oaie/restaurant2-restaurant2) should be removed from the gold-
standard because they do not refer to the same restaurant. There descriptions are as
follows:
(1) [’name: hotel bel air’, ’category: californian’, ’phone number: 310/472-1211’ , has-
address[street:’701 stone canyon rd.’, is in city[name:bel air’]]]
(2) [’name: art’s deli’, has category[’name:delis’], ’phone number:818-762-1221’ ,
has-address[’city:studio city’, street:’12224 ventura blvd.’]]
3 General comments
3.1 Comments on the results
The main strength of our system is its capacity to ensure a good precision in the re-
sults. In the person data set, it shows its strength over ontology knowledge reasoning
and similarity measures adaptation. LN2R system is also able to minimize the num-
ber of comparisons thanks to the filtering step which leads to improvement in running
time. The weak points are: the absence of knowledge on the functionality of proper-
ties impacts the performance of the system. LN2R system works on data sets which
should conform to the same ontology are for which the ontology alignment is already
performed.
177
3.2 Discussions on the way to improve the proposed system
Our system may be improved by several ways. We are studying how LN2R can be
extended to take into account alignments between classes and properties. We also want
to optimize the system in order to insure its scalability.
3.3 Comments on the OAEI 2010 test cases
It will be interesting to provide test cases where the alignment inference is global. It
means that the alignments may concern several kinds of entities e.g. persons, addresses,
books, etc. It will be useful also to have data sets which conform to the same ontology
or at least give the alignments between their corresponding ontologies.
4 Conclusion
Instance matching is very important to realize the semantic Web ambitions by facili-
tating interoperability of ontology based applications. In this paper, we have presented
the promising results of LN2R system for its first participation in the instance match-
ing track of OAEI 2010. By this experience, we have shown LN2R strengths when the
ontology knowledge is rich. In the person data sets, LN2R has obtained very good re-
sults and reasonable ones for the restaurant data sets. As future work, we will study the
extension of LN2R to the general problem of matching ontologies with instances. We
also plan to optimize LN2R by designing a distributed inference algorithm.
References
1. Shvaiko, P., Euzenat, J.: A survey of schema-based matching approaches. (2005) 146–171
2. Rahm, E., Bernstein, P.A.: A survey of approaches to automatic schema matching. The VLDB
Journal 10(4) (2001) 334–350
3. Batini, C., Scannapieco, M.: Data Quality: Concepts, Methodologies and Techniques (Data-
Centric Systems and Applications). Springer-Verlag New York, Inc., Secaucus, NJ, USA
(2006)
4. Elmagarmid, A.K., Ipeirotis, P.G., Verykios, V.S.: Duplicate record detection: A survey. IEEE
Trans. on Knowl. and Data Eng. 19(1) (2007) 1–16
5. Saıs, F., Pernelle, N., Rousset, M.C.: Combining a logical and a numerical method for data
reconciliation. J. Data Semantics 12 (2009) 66–94
6. Dong, X., Halevy, A.Y., Madhavan, J.: Reference reconciliation in complex information
spaces. In: SIGMOD Conference. (2005) 85–96
7. Saıs, F., Pernelle, N., Rousset, M.C.: L2r: A logical method for reference reconciliation. In:
AAAI. (2007) 329–334
8. Golub, G.H., Loan, C.F.V.: Matrix computations (3rd ed.). Johns Hopkins University Press,
Baltimore, MD, USA (1996)
9. Cohen, W.W., Ravikumar, P.D., Fienberg, S.E.: A comparison of string distance metrics for
name-matching tasks. In: IIWeb. (2003) 73–78
178
MapPSO Results for OAEI 2010
Jurgen Bock1
FZI Forschungszentrum Informatik, Karlsruhe, [email protected]
Abstract. This paper presents and discusses the results produced bythe MapPSO system for the 2010 Ontology Alignment Evaluation Ini-tiative (OAEI). MapPSO is an ontology alignment approach based ondiscrete particle swarm optimisation (DPSO). Firstly, specific character-istics of the MapPSO system and their relation to the results obtainedin the OAEI are discussed. Secondly, the results for the benchmarks anddirectory tracks are presented and discussed.
1 Presentation of the system
With the 2008 OAEI campaign the MapPSO system (Ontology Mapping byParticle Swarm Optimisation) was introduced [1] as a novel approach to tacklethe ontology alignment problem by applying the technique of particle swarmoptimisation (PSO).
1.1 State, purpose, general statement
The development of the MapPSO algorithm has been motivated by the followingobservations:
1. Ontologies are becoming numerous in number and large in size.2. Ontologies evolve gradually.3. Ontologies differ in key characteristics that can be exploited in order to
compute alignments.
Solving the ontology alignment problem using a PSO-based approach, as doneby the MapPSO system, tackles these observations as follows:
1. PSO works inherently parallel, such that large ontologies can be aligned ona parallel computation infrastructure.
2. PSO works incrementally, which allows the algorithm to start with an initialor partial configuration (i.e. for instance an alignment of previous ontologyversions) and refine it as the ontologies evolve.
3. PSO works as a meta-heuristic, i.e. independently of the objective functionto be optimised. In the case of ontology alignment this means that the ob-jective function can be adjusted according the particular alignment scenarioat hand.
179
The idea of the MapPSO approach is to provide an algorithm that fulfils theaforementioned characteristics. Particularly the focus is not to provide a univer-sal library of similarity measures (base matchers) to form that specific objectivefunction to be optimised, but rather to provide a scalable mechanism that canused with various objective functions depending on the alignment scenario athand.
MapPSO is still in the status of a research prototype, where recent workhas been done exploiting the parallel nature of the algorithm in a cloud-basedinfrastructure [2].
1.2 Specific techniques used
MapPSO treats the ontology alignment problem as an optimisation problemand solves it by applying a discrete particle swarm optimisation (DPSO) algo-rithm [3]. To this end, each particle in the swarm represents a valid candidatealignment, which is updated in an iterative fashion. In each iteration, knowingabout the particle representing the best alignment in the swarm, other particlesadjust their alignments, influenced by this best particle. A random componentwhen adjusting an alignment makes sure that the swarm does not converge to alocal optimum.
In MapPSO the quality of an alignment is determined by the average of thequalities of its correspondences, as well as by the number of correspondences inthe alignment1. Each correspondence is evaluated by a number of base matchers,whose evaluation values are aggregated by a specified aggregator. Base matchersand aggregator can be selected via the params.xml configuration file. This mech-anism makes MapPSO highly adjustable, since different alignment scenarios willmost likely require different base matchers in order to determine similarity be-tween entities. By following the instructions in the MapPSO documentation2
one can easily develop base matchers and aggregators tailored to a particularalignment scenario at hand.
1.3 Adaptations made for the evaluation
Some OAEI tracks do not evaluate correspondences of all entity types. For in-stance in the benchmarks track, no instance correspondences are part of thereference alignments, while in the instance matching track only instance corre-spondences are part of the reference alignments. For this reason, an additionalparameter was introduced for the MapPSO command-line interface that allows
1 Apart from striving for correct correspondences, it is necessary to identify the cor-rect number of correspondences, which is done in MapPSO by preferring largeralignments to smaller ones.
2 http://sourceforge.net/apps/mediawiki/mappso/index.php?title=Guide_for_
implementing_base_matchers and http://sourceforge.net/apps/mediawiki/
mappso/index.php?title=Guide_for_implementing_aggregation_functions
respectively.
180
the user to specify which correspondence types are to be included in the producedalignment.
1.4 Link to the system and parameters file
The release of MapPSO (MapPSO.jar) and the parameter file params.xml usedfor OAEI 2010 are located in MapPSO.zip at http://sourceforge.net/projects/mappso/files/ in the folder oaei2010.
1.5 Link to the set of provided alignments (in align format)
The alignments of the OAEI 2010 as provided by MapPSO are located in thefile alignments.zip at https://sourceforge.net/projects/mappso/files/in the folder oaei2010.
2 Results
The benchmarks track was via the (preliminary) SEALS platform3. To this endMapPSO has been provided as a web service4 For the directory track, resultswere computed offline and sent to the track organiser.
2.1 benchmark
As from last year’s participation it became apparent that MapPSO performsbetter with respect to relaxed precision and recall measures than with respectto classical measures [4, 5]. For this reason, the symmetric precision and recallmeasures were computed5 in addition to the classical measures as provided bythe SEALS platform. Figures 1 and 2 illustrate classical and symmetric precision,and classical and symmetric recall respectively. It shall be noted that MapPSOis a non-deterministic method and therefore on a set of independent runs thequality of the results and the number of mappings in the alignments will besubject to slight fluctuations. The plots in Figures 1 and 2 were generated ina different run, than in the results obtained using the SEALS platform, thusresults might not match completely.
The reason for MapPSO performing significantly better in terms of symmetricprecision and recall is due to the fact that the algorithm keeps good correspon-dences not allowing them to be discarded in a later iteration. This, however,prevents entities participating in such good correspondences to participate in aneven better correspondence in a later iteration. In case this better correspondencewould be the correct one with respect to the reference alignment, the good onefound is counted as wrong with respect to classical evaluation metrics, while itscloseness is respected in the relaxed metrics.
3 http://seals.inrialpes.fr/platform/4 Web service end point: http://krake16.perimeter.fzi.de:8080/MapPSOWS5 Relaxed precision and recall measures were computed using the methods providedby the Alignment API.
181
0 0.2 0.4 0.6 0.8 1
101103104201
201-2201-4201-6201-8
202202-2202-4202-6202-8
203204205206207208209210221222223224225228230231232233236237238239240241246247248
248-2248-4248-6248-8
249249-2249-4249-6249-8
250250-2250-4250-6250-8
251251-2251-4251-6251-8
252252-2252-4252-6252-8
253253-2253-4253-6253-8
254254-2254-4254-6254-8
257257-2257-4257-6257-8
258258-2258-4258-6258-8
259259-2259-4259-6259-8
260260-2260-4260-6260-8
261261-2261-4261-6261-8
262262-2262-4262-6262-8
265266301302303304
Sym
metric P
recisionC
lassical Precision
Fig. 1. Classical vs. Symmetric Precision from OAEI 2010.
182
0 0.2 0.4 0.6 0.8 1
101103104201
201-2201-4201-6201-8
202202-2202-4202-6202-8
203204205206207208209210221222223224225228230231232233236237238239240241246247248
248-2248-4248-6248-8
249249-2249-4249-6249-8
250250-2250-4250-6250-8
251251-2251-4251-6251-8
252252-2252-4252-6252-8
253253-2253-4253-6253-8
254254-2254-4254-6254-8
257257-2257-4257-6257-8
258258-2258-4258-6258-8
259259-2259-4259-6259-8
260260-2260-4260-6260-8
261261-2261-4261-6261-8
262262-2262-4262-6262-8
265266301302303304
Sym
metric R
ecallC
lassical Recall
Fig. 2. Classical vs. Symmetric Recall from OAEI 2010.
183
2.2 directory
In the directory track the same set of parameters was used as in the other tracks,which includes the same set of base matchers. Due to the nature of the datasetsin this track, several base matchers were not applicable or might even havecontributed in a counterproductive way. For instance in the singletask subtrack,meaningless IDs were used as URI fragments, which could lead to a high similar-ity of those entities with a similar ID with respect to a particular base matcher.Since this information is known before running the matcher, deactivating thisbase matcher might have lead to better results. Additionally, MapPSO does notfilter final results according to the confidence values gained. Thus, in its cur-rent implementation, many bad correspondences are left in the final alignment,reducing precision.
3 General comments
In the following some general statements about the OAEI procedure, modalities,and results obtained are given.
3.1 Comments on the results
Compared to the results of the 2009 benchmarks track, a slight decrease in thesymmetric precision and recall measures can be observed. This is due to the factthat this year, MapPSO has been configured with a stronger focus on finding thecorrect size of an alignment. The configuration used in 2009 was rather tailoredto the benchmarks track where the alignment is known to contain all entities.Disregarding this assumption causes the symmetric recall measure to drop, butmakes the system more suitable for real-world use cases.
3.2 Discussions on the way to improve the proposed system
MapPSO is currently being worked on in order to incorporate a guided searchcomponent for two reasons. Firstly, it is expected to increase convergence speed,and secondly it is expected to improve classical precision and recall due to thereasons explained in Sect. 2.1.
3.3 Comments on the OAEI 2010 procedure
The OAEI modalities require participating systems to use the same parameterconfiguration for each track and each test case. According to assumption 3 statedin Sect. 1.1 different alignment scenarios will most likely require different meansof determining a good alignment. Assuming that an alignment tool will not usedin an out-of-the-box configuration in any real-world alignment task, makes thisrequirement of a single (and thus compromised) parameter configuration ratherartificial.
184
4 Conclusion
The MapPSO system was described briefly with respect to the idea behind itsDPSO-based approach. The results obtained by the MapPSO system for theOAEI 2010 tracks benchmarksand directory were presented. Several observationsregarding these results were highlighted, in particular the significant differencebetween classical and symmetric precision and recall. Also the effect of having asingle configuration throughout all OAEI tracks were discussed.
Future development of MapPSO will be targeted towards user interaction aswell as alignment refinement. As for the latter, an initial (partial) alignment,such as a previous version of an alignment, can be given as a start configurationof a particle, which is then refined by running the algorithm.
Acknowledgements
The author would like to thank Michael Mutter and Carsten Danschel for theirvaluable contributions and feedback.
References
1. Bock, J., Hettenhausen, J.: MapPSO Results for OAEI 2008. In Shvaiko, P., Eu-zenat, J., Giunchiglia, F., Stuckenschmidt, H., eds.: Proceedings of the 3rd Interna-tional Workshop on Ontology Matching (OM-2008). Volume 431 of CEURWorkshopSeries., CEUR-WS.org (November 2008)
2. Bock, J., Lenk, A., Danschel, C.: Ontology Alignment in the Cloud. In: Proceedingsof the 5th International Workshop on Ontology Matching (OM-2010), http://ceur-ws.org, CEUR Workshop Proceedings (November 2010)
3. Bock, J., Hettenhausen, J.: Discrete Particle Swarm Optimisation for OntologyAlignment. Information Sciences Article in Press (2010)
4. Ehrig, M., Euzenat, J.: Relaxed Precision and Recall for Ontology Matching. InAshpole, B., Ehrig, M., Euzenat, J., Stuckenschmidt, H., eds.: Proceedings of theK-CAP 2005 Workshop on Integrating Ontologies. Volume 156 of CEUR WorkshopProceedings., CEUR-WS.org (October 2005) 25–32
5. Euzenat, J., Ferrara, A., Hollink, L., Isaac, A., Joslyn, C., Malaise, V., Meilicke,C., Nikolov, A., Pane, J., Sabou, M., Scharffe, F., Shvaiko, P., Spiliopoulos, V.,Stuckenschmidt, H., Svab-Zamazal, O., Svatek, V., dos Santos, C.T., Vouros, G.,Wang, S.: Results of the Ontology Alignment Evaluation Initiative 2009. In Shvaiko,P., Euzenat, J., Giunchiglia, F., Stuckenschmidt, H., Noy, N., Rosenthal, A., eds.:Proceedings of the 4th International Workshop on Ontology Matching (OM-2009).Volume 551 of CEUR Workshop Proceedings., http://ceur-ws.org (October 2009)
185
Results of NBJLM for OAEI 2010
Song Wang1,2, Gang Wang1 and Xiaoguang Liu1
1College of Information Technical Science, Nankai University Nankai-Baidu Joint Lab,
Weijin Road 94, Tianjin, China2Military Transportation University, The Equipment support Department, Tianjin, China
Abstract. This paper presents the results obtained by NBJLM (Nankai Baidu
Joint Lab Matcher) for its first participation to OAEI 2010. The research of
ontology-based similarity calculation among concepts has already been a hot is-
sue. NBJLM is an hybrid ontology alignment method that considers both simi-
larity of literal concept and semantic structure. Simultaneously, how to accelerate
matching has been mentioned in this paper and the experimental results show the
remarkable improvement of matching speed. In OAEI 2010, NBJLM submitted
the result for one alignment task: anatomy.
1 Presentation of NBJLM
In recent years, Ontology matching is mainly used in ontology integration, ontology
merging, and ontology reusing. Many approaches to ontology matching have been pro-
posed over the years, references[1][5][3] make full use of information, probability and
statistics theory, however, they have limited ability to distinguish semantic differences,
and the similarity calculation methods are not perfect. Besides, references[2][4][7][8]
have considered various factors, but they do not take into account how to avoid unneces-
sary calculation to shorten computing time in mapping large-scale ontologies. NBJLM
is a multiple strategy dynamic ontology matching system implemented in java. It con-
siders both the literal concept and ontology structure that includes node depth, node
density and semantic distance.
1.1 State, purpose, general statement
Given two heterogeneous ontologies O1 and O2, a matching is made up of a set of cor-
respondences between pairs of node IDs belonging to O1 and O2, respectively. NBJLM
is designed to find out relations of equivalence and subsumption between entities, i.e.
classes and properties, issued from two ontologies. Our approach makes use of the
matching strategy that considers literal similarity measure and ontology structure sim-
ilarity measure. The core contributions of NBJLM is described as followed: Firstly, it
uses Hash mapping algorithm to improve efficiency of calculation. Secondly, it takes a
full analysis of a number of issues to be considered in structure matching, which makes
the algorithm works better, and the matching results are more accurate and efficient. As
demonstrated by the experimental results, our method can greatly cut the running time,
meanwhile, precise matching results can be obtained.
186
WordNet
HeterogeneousOntologies
Structure SimilarityMatch Results
1 1
2 23 3
1 123 2
347
4
5
Literal Similarity
Fig. 1. Procedure of the matching of heterogeneous ontologies.
1.2 Specific techniques used for Anatomy Track
NBJLM uses a new matching strategy that considers literal similarity measure and on-
tology structure similarity, simultaneously. We obtain the following formula:
𝑆𝑖𝑚(𝐼𝐷1, 𝐼𝐷2) = 𝜃 × 𝑆𝑖𝑚 𝑙𝑖𝑡𝑒𝑟𝑎𝑙(𝐼𝐷1, 𝐼𝐷2) + (1− 𝜃)× 𝑆𝑖𝑚 𝑠𝑡𝑟𝑢𝑐𝑡(𝐼𝐷1, 𝐼𝐷2)
where 𝑆𝑖𝑚 𝑙𝑖𝑡𝑒𝑟𝑎𝑙(𝐼𝐷1, 𝐼𝐷2) is the literal concept similarity measure, 𝑆𝑖𝑚 𝑠𝑡𝑟𝑢-
𝑐𝑡(𝐼𝐷1, 𝐼𝐷2) is the structural similarity measure, and 𝜃 (0 < 𝜃 < 1) is paramater to
control how much literal and ontology structure contribute to the ontologies matching
respectively. Firstly, the measure of literal similarity is a preliminary matching. It takes
account of polysemy and synonym of a word, by transforming the word into a semantic
collection using WordNet. Then we can get the preliminary matching results that is se-
mantic mapping rather than spelling mapping of words. Secondly, based on the literal
matching results, the measure of ontology structure similarity is calculated through the
relation between hypernym and hyponym of a word, considering distance of edges, and
depth and density of node in the hierarchy of ontology. With the final combination of
the two values, and with adjustment of the parameter, we could obtain more reasonable
matching results. The procedure is shown in Fig. 1.
An optimized algorithm for concept sets retrieving If look up a word in Word-
Net, we can get one or more Synsets (defined by WordNet). For one thing each Synset
is a concept set of the words which have the same meaning. For another a word may
have several meanings, therefore, each Synset can be used to express one concept of
the word. The concept of a node ID in the hierarchy of ontology may be described by
several phrases, which are composed of words. That means the concept of the node ID
could be described by several Synsets. If we deal with all the Synsets in matching, re-
dundant computation will be inevitable. Therefore, this paper proposes a strategy that
obtain the set of Synsets, which are the most similar to the concept of the phrase while
187
a<ba>b d=e d<ed>e
Phrase p
A
a=b
S3
WordNet
S4 S5S2S1
S3’ S4’ S5’S2’S1’
Add hypernym and hyponym
W1
S2’’ S3’’ S4’’ S5’’
W2 W3
S1’’
W1
S1S2
W2
S3
W3
S4
W3
S4S5
W3
S5
W1
S1
W1
S2
S1S2S3S4S5
S1 S3 S4 S5
S1S3S4S5
B
a b c d e
Fig. 2. Get the optimal Synsets to describe the concept of a phrase
include the least Synsets, to describe the concept of a node ID with the help of Word-
Net. So the unnecessary computation work could be reduced. Fig. 2 describes a simple
example that how to tackle a phrase to get the optimal Synsets:
1) Obtain the Synsets of all the words 𝑤1, 𝑤2 and 𝑤3 from phrase 𝑝 got from a node
of 𝑂1(or 𝑂2) by WordNet, respectively, (𝑆1, 𝑆2), (𝑆3) and (𝑆4, 𝑆5).2) Get the union set of all the Synsets, 𝐴(𝑆1, 𝑆2, 𝑆3, 𝑆4, 𝑆5), which denotes the con-
cept of phrase p and includes the most Synsets.
3) Add the semantic environment (hypernym and hyponym of the Synset) to Synset.
Then we get 𝑆1′, 𝑆2′, 𝑆3′, 𝑆4′ and 𝑆5′.4) Intersect 𝐴(𝑆1, 𝑆2, 𝑆3, 𝑆4, 𝑆5) with 𝑆1′, 𝑆2′, 𝑆3′, 𝑆4′ and 𝑆5′, separately, result-
ing in 𝑆1′′, 𝑆2′′, 𝑆3′′, 𝑆4′′ and 𝑆5′′. And the numbers of elements of the inter-
sections are 𝑎, 𝑏, 𝑐, 𝑑 and 𝑒. Meanwhile, establish correspondences between 𝑆1 and
𝑆1′′, 𝑆2 and 𝑆2′′, 𝑆3 and 𝑆3′′, 𝑆4 and 𝑆4′′, 𝑆5 and 𝑆5′′. The purpose of doing in-
tersections is to find correlation between semantic environment of a Synset and the
concept of phrase p. The larger number of the intersection’s elements is, the more
similar relationship between them is.
5) Compare the numbers of intersections’ elements mentioned at step 4, which are
generated from the same word. And select the Synset of each word, associated with
the result of intersection which has the larger number of elements. For example, on
the assumption that 𝑎 > 𝑏, 𝑑 = 𝑒 (c has no comparable object), the Synsets of 𝑤1,
𝑤2 and 𝑤3 are (𝑆1), (𝑆3) and (𝑆4, 𝑆5), respectively.
6) Get the union set of (𝑆1), (𝑆3) and (𝑆4, 𝑆5), 𝐵(𝑆1, 𝑆3, 𝑆4, 𝑆5), which denotes
the concept of phrase p.
188
7) It can be found that Synset 𝑆2 existing in 𝐴 but not in 𝐵 is uncorrelated to the
concept of phrase 𝑝. Therefore, the redundancy can be filtered out by our optimized
algorithm. Besides, as increasing in the number of words of phrase, the optimization
of the algorithm could be more obvious. Since the matching of nodes in the ontolo-
gies is based on the matching of Synsets, the reduction of Synsets, which denote the
concepts of nodes in the ontology, will inevitably lead to the reduction of irrelevant
semantic mappings and greatly reduce the amount of calculation.
Method of calculation of structural similarity The calculation of structural similar-
ity involves semantic distance with weight, information content, depth and density of
node. In order to tackle two ontologies conveniently, we add a virtual common root
node which connects two ontologies. So the model could be changed from two inde-
pendent ontologies to a large ontology, which facilitates the matching. The process of
matching is described as follow: firstly, search the common ancestor 𝐶 of two nodes
𝑐1 and 𝑐2. In fact, 𝐶 is a mapping pair(𝑐1′, 𝑐2′) got from the matching results of lit-
eral concepts, where 𝑐1′ is the ancestral node of 𝑐1 and 𝑐2′ is the ancestral node of 𝑐2.
Secondly, calculate the semantic distance between 𝑐1 and 𝑐2 through 𝐶. Thirdly, do
iterative calculation that search the common ancestor 𝐶 of 𝑐1′ and 𝑐2′ until 𝐶 is the
virtual common node. Finally, add depth and density of nodes into the calculation. The
formula is:
𝑆𝑖𝑚 𝑠𝑡𝑟𝑢𝑐𝑡(𝐼𝐷1, 𝐼𝐷2) = 𝑆𝑖𝑚(𝐶𝑜𝑚 𝑎𝑛𝑐𝑒𝑠𝑡𝑜𝑟(𝑐1, 𝑐2)
)× 𝛼𝑘
𝑘 +𝐷𝑖𝑠(𝑐1, 𝑐2)+
𝛽(𝜂 + (1− 𝜂)× 𝑒(𝑐1) + 𝑒(𝑐2)
2
)+𝛾
2
( 𝑑(𝑐1)
𝑑(𝑐1) + 1+
𝑑(𝑐2)
𝑑(𝑐2) + 1
)
Where 𝐶𝑜𝑚 𝑎𝑛𝑐𝑒𝑠𝑡𝑜𝑟(𝑐1, 𝑐2) returns the common ancestor pair of 𝑐1 and 𝑐2, and
𝐷𝑖𝑠(𝑐1, 𝑐2) is the semantic distance, 𝑒(𝑐1) and 𝑑(𝑐1) are the density and depth of
node[7]. The parameters 𝑘(𝑘 > 0), 𝜂(0 < 𝜂 < 1), 𝛼, 𝛽 and 𝛾 (𝛼+ 𝛽 + 𝛾 = 1) control
how much semantic distance, depth, density contribute to the calculation of structural
similarity respectively.
𝐷𝑖𝑠(𝑐1, 𝑐2) =∑
𝑥∈𝑝𝑛(𝑐1)
𝑤𝑡(𝑥, 𝑝(𝑥)
)+
∑𝑥∈𝑝𝑛(𝑐2)
𝑤𝑡(𝑝(𝑥), 𝑥
)
𝑤𝑡(𝑐, 𝑥) = 𝐿𝑠(𝑐, 𝑥)× 𝑇 (𝑐, 𝑥)
𝐿𝑠(𝑐, 𝑥) = − log(𝑃 (𝑐∣𝑥)
)= − log
𝑃 (𝑐∩𝑥)
𝑃 (𝑥)= 𝐼𝐶(𝑐)− 𝐼𝐶(𝑥)
Where 𝑤𝑡(𝑐, 𝑥) is the weight of 𝑒𝑑𝑔𝑒(𝑐, 𝑥), 𝑝𝑛(𝑐) is the set of nodes which are
on the path from node 𝑐 to the common ancestor node, 𝑝(𝑥) is the parent node of 𝑥,
𝐼𝐶(𝑥)is interest degree[6], 𝐿𝑠(𝑐, 𝑥) is the difference of the information content values
between a child node and its parent, and 𝑇 (𝑐, 𝑥) is the link relation factor.
There is something important to pay attention to, which makes the algorithm more
efficiency:
189
a b c kO1:ID-1
a b d k O2:ID-1'
a b c e O2:ID-2'
Fig. 3. Literal concept mapping of one to many
– This approach searches all the ancestor nodes of two nodes to be matched, and
select the best matching path. If only search the nearest common ancestor node, the
result may be wrong. For example: owing to the situation of one to many mappings
in the matching results of literal concepts, it may occur that the mappings (𝑂1 :𝐼𝐷 − 2, 𝑂2 : 𝐼𝐷 − 2′) and (𝑂1 : 𝐼𝐷 − 2, 𝑂2 : 𝐼𝐷 − 4′) got from results
of literal concept matching are candidates for structural matching, but in fact (𝑂1 :𝐼𝐷−2, 𝑂2 : 𝐼𝐷−2′) is the best mapping. When comparing the node𝑂1 : 𝐼𝐷−6and node 𝑂2 : 𝐼𝐷 − 6′, if only search their nearest common ancestor, we will get
a pair of nodes, 𝑂1 : 𝐼𝐷− 2 and 𝑂2 : 𝐼𝐷− 4. However, it is not the best mapping
pair (we have known that the pair of 𝑂1 : 𝐼𝐷− 2 and 𝑂2 : 𝐼𝐷− 2′ is the best). To
avoid this, we need to traverse all the common ancestors of nodes rather than the
nearest. Then compare the iterative results and choose the best.
– Involve the literal interest degree. For instance, when we find mapping pairs (𝑂1 :𝐼𝐷 − 1, 𝑂2 : 𝐼𝐷 − 1′) and (𝑂1 : 𝐼𝐷 − 1, 𝑂2 : 𝐼𝐷 − 2′) have the same struc-
tural similarity, and the values of their literal similarity calculations are both 3/4 as
shown in fig. 3, where 𝑎, 𝑏, 𝑐, 𝑑, 𝑒 and 𝑘 are Synsets, then the literal interest degree is
needed to judge which the better matching object of𝑂1 : 𝐼𝐷−1 from𝑂2 : 𝐼𝐷−1′
and 𝑂2 : 𝐼𝐷 − 2′ is: the less frequency of a Synset occurs in the ontology is, the
more it contributes to the meaning of the node. So we calculate all the literal interest
degrees of the common Synsets in each mapping pair using the formula metioned
in Definition 4. And compare the maximal literal interest degrees of all the mapping
pairs, then the max is the best matching because they contain the common Synset
whose meaning is closer to concept of the phrase. To suppose the maximal literal
interest degree of (𝑂1 : 𝐼𝐷 − 1, 𝑂2 : 𝐼𝐷 − 1′) is 𝑛1 got from 𝑘, simultaneously,
the maximal literal interest degree of (𝑂1 : 𝐼𝐷 − 1, 𝑂2 : 𝐼𝐷 − 1′) is 𝑛2 got from
𝑒, and 𝑛1 > 𝑛2, we can draw the conclusion: (𝑂1 : 𝐼𝐷− 1, 𝑂2 : 𝐼𝐷− 1′) should
be the best mapping pair because 𝑂1 : 𝐼𝐷 − 1 is more interested in Synset 𝑘.
– At last calculate the factors of density and depth of node. Because in each iteration
the value of semantic distance should be multiplied by similarity of the common
ancestor node which is smaller than 1, it will surely lead to the similarity of child
nodes smaller than those of their ancestor nodes. This is contradictory to the role of
depth and density calculation, because the nodes which have greater values of depth
and density will have the larger value of similarity. Therefore, we must calculate the
depth and density of node out of the procedure of calculation of semantic distance
and iterations.
190
Read O1
WordNetFor O1
Main Thread
Read O2
WordNetFor O2
Hash O1
New Thread_1
New Thread_N
...
Literal Concept & Structure Matching
...
Main Thread
New Thread_N
Return
New Thread_1Synchrony
Fig. 4. Parallelization of the algorithm implemented by multi-threads
Parallelization of the algorithm NBJLM uses parallel algorithm to accelerate the
process of matching. Fig. 4 shows the task partitioning. Firstly, we use the main thread
to read 𝑂1 file and then look up the Synsets of all the node IDs of 𝑂1 in the WordNet.
The reason of use only one thread is that this stage contains only IO operations which
can not benefit from parallel execution and WordNet does not provide thread-safe APIs.
Secondly, another multi-threads are launched to calculate hash values of node IDs’
Synsets of 𝑂1, meanwhile we use the main thread to read 𝑂2 file and look up the
Synsets of all the node IDs of 𝑂2. And these tasks could be run in parallel because
one part is CPU operation, and another is IO operation. Finally, we synchronize all the
threads, and then use them to calculate the literal concepts similarity and the structure
similarity.
1.3 Adaptations made for the evaluation
This year, NBJLM has first taken part in OAEI. Therefore, in OAEI 2010 NBJLM used
the match to compute the alignments for one track(anatomy). In order to assure the
matching process is fully automated, all parameters are configured automatically with
a strategy. No specific adaptations have been made.
1.4 Link to the system and parameters file
The version of NBJLM for OAEI 2010 can be downloaded from our website: ℎ𝑡𝑡𝑝 ://𝑤𝑤𝑤.𝑏𝑟𝑠𝑏𝑜𝑥.𝑐𝑜𝑚/𝑂𝐴𝐸𝐼2010. The parameter file is also included in the NBJLM.zip
191
file. I recommend readers to read the readme.txt file first. The file includes the necessary
description and parameters as well in brief.
1.5 Link to the set of provided alignments (in align format)
NBJLM alignment results for OAEI can be found at
ℎ𝑡𝑡𝑝 : //𝑤𝑤𝑤.𝑏𝑟𝑠𝑏𝑜𝑥.𝑐𝑜𝑚/𝑂𝐴𝐸𝐼2010.
2 Results
In this section, we describe the results of NBJLM algorithm against the Anatomy on-
tologies provided by the OAEI 2010 campaign. In this test, the real world cases of
anatomy for Adult Mouse Anatomy (2744 classes) and NCI Thesaurus (3304 classes)
for human anatomy are included. This year we have participated in task#1 for the first
time. Experiments were done on a computer with 1.8GHz AMDAthlon dual-core CPU
and 2GB DDR2 RAM memory.
2.1 anatomy
Subtrack#1 In this subtrack, participants are asked to maximize F-measure. NBJLM
used a threshold equal to 0.8 and obtained an F-measure equal to 85.8%. NBJLM ob-
tained precision equal to 92.0% and recall equal to 80.3%. The runtime was 2 minutes.
3 General comments
3.1 Comments on the results
– Strengths NBJLM deals with ontology from two different views and combines
results of every step in sequential way. If the ontologies have regular literals and
hierarchical structures, NBJLM can achieve satisfactory alignments. And the way
of minimizing the comparisons between entities, which leads to enhance running
efficiency.
– Weaknesses NBJLM depends on the literal concept results to calculate structural
similarity. So if the literals of concept missed, NBJLM will get bad results.
3.2 Discussions on the way to improve the proposed system
1) To enrich the semantic dictionaries because WordNet which is not a professional
dictionary cannot obtain more comprehensive semantic concepts.
2) To take into account all concepts properties instead of only the hierarchicals ones.
192
4 Conclusion
This paper reports our first participation in OAEI campaign. We present the alignment
process of NBJLM and describe the specific techniques for ontology matching. The
method based on heterogeneous ontologies combines the calculations of literal con-
cept and ontology structure and pays more attention to computational efficiency. The
strengths and the weaknesses of our proposed approach are summarized and the pos-
sible improvement will be made for the system in the future. We propose a brand new
algorithm to match ontologies.
References
1. E. Agirre and G. Rigau. A proposal for word sense disambiguation using conceptual distance.
AMSTERDAM STUDIES IN THE THEORY AND HISTORY OF LINGUISTIC SCIENCE SE-RIES 4, pages 161–172, 1997.
2. Y. Jean-Mary, E. Shironoshita, and M. Kabuka. Ontology matching with semantic verification.
Web Semantics: Science, Services and Agents on the World Wide Web, page 235C251, 2009.
3. Y. Li, D. McLean, Z. Bandar, J. O’Shea, and K. Crockett. Sentence similarity based on se-
mantic nets and corpus statistics. IEEE Transactions on Knowledge and Data Engineering,
18(8):1138–1150, 2006.
4. M. A. Q. Muhammad Fahad. Similarity Computation by Ontology Merging System: DKP-
OM. Computer, Control and Communication, pages 17–18, February 2009.
5. P. Resnik et al. Using information content to evaluate semantic similarity in a taxonomy. In
International Joint Conference on Artificial Intelligence, volume 14, pages 448–453. Citeseer,
1995.
6. S. Ross. A first course in probability. New York, 1994.
7. J. Sevilla, V. Segura, A. Podhorski, E. Guruceaga, J. Mato, L. Martinez-Cruz, F. Corrales,
and A. Rubio. Correlation between gene expression and GO semantic similarity. IEEE/ACMTransactions on Computational Biology and Bioinformatics (TCBB), 2(4):338, 2005.
8. J. Tang, J. Li, B. Liang, X. Huang, Y. Li, and K. Wang. Using Bayesian decision for ontology
mapping. Journal of Web Semantics: Science, Services and Agents on the WorldWideWeb,
pages 243–262, 2006.
193
RiMOM Results for OAEI 2010
Zhichun Wang1, Xiao Zhang1, Lei Hou1,Yue Zhao2, Juanzi Li1, Yu Qi3 , Jie Tang1
1Tsinghua University, Beijing, China {zcwang,zhangxiao,greener,ljz,tangjie}@keg.cs.tsinghua.edu.cn
2Beihang University, Beijing, China [email protected]
3National University of Defense Technology, Changsha, China [email protected]
Abstract. This paper presents the results of RiMOM in the Ontology Alignment Evaluation Initiative (OAEI) 2010. We participate in three tracks of the campaign: Benchmark, IM@OAEI2010 (IMEI), and Very Large Crosslingual Resources (VLCR). We first describe the basic alignment process and alignment strategies in RiMOM, and then we present specific techniques used for different tracks. At last we give some comments on our results and discuss some future work on RiMOM.
1 Presentation of the system
Recently, ontology alignment has been developed as a key technology to solve interoperability problems across heterogonous data sources. Many automatic ontology alignment systems have been proposed and achieve good performance in real world data. With the development of Linked Data [1] and various social network websites, huge amount of semantic data is published on the web, which not only poses new challenges over traditional schema level ontology alignment algorithms, but also demands new techniques for instance matching.
RiMOM is a multistrategy dynamic ontology alignment system [2]. It implements several different matching strategies which are defined based on different ontological information. For each individual matching task, RiMOM can automatically and dynamically combine multiple strategies to generate a composed matching result.Recently, some new features were added into the new version of RiMOM which enable it to deal with unbalanced ontology matching [3], user interactive ontology matching [4], and large scale instance matching.
1.1 State, purpose, general statement
Currently, RiMOM is developed with a flexible framework for ontology alignment, where different kinds of alignment strategies can be plugged and configured easily. Fig 1 shows the architecture of RiMOM system.
The whole system consists of three layers: interface layer, task layer and
194
component layer. In the interface layer, RiMOM provides a graphical user interface to allow users to customize the matching procedure: including selecting preferred components, setting the parameters for the system, etc. In semi-automatic ontology matching, user can also get involved in the matching process via the user interface. The task layer stores parameters of the alignment tasks, and controls the execution process of components in the component layer. In component layer, we define five groups of executable components, including preprocessor, matcher, aggregator, postprocessor and evaluator. In each group, there are several instantiated components. For a certain alignment task, user can select appropriate components and execute them in desired sequence.
JenaPreprocessor
OW
LAPIPreprocessor
SimilarityFlooding
Preprocessor
GaussianFunctionPostprocessor
ThresholdFilter
SimilarityFlooding
Matchor
EditdistanceMatchor
WordN
etMatchor
VectorB
asedMatchor
MachineLearningB
asedM
atchor
AverageW
eightedA
ggregator
ConsistenceW
eightedA
ggregator
SigmoidW
eightedA
ggregator
IndirectMatching
Aggregator
PRFEvaluator
Fig 1. Architecture of RiMOM system
1.2 Specific techniques used
This year we participate in three tracks of the campaign: Benchmark, IM@OAEI2010 (IMEI), and Very Large Crosslingual Resources (VLCR). We describe specific techniques used in different tracks as follows: Benchmark Track
For benchmark track, we use three matching strategies: (1) Name based strategy: In this strategy, we calculate the edit distance between
labels of two entities. Edit distance estimates the number of operation needed to convert one string into another. We define 1 2(1 # / ( , ))op max_length l l as the similarity of two labels, where #op indicates the number of operations,
1 2( , )max_length l l represents the maximal length of the two labels. (2) Metadata based strategy: In this strategy, we treat the information of each
entity as a document, which consists of words in entity’s label and comment. Then we construct a weighted feature vector using tf-idf technology, the similarity between two entities is then calculated as the cosine of the two vectors.
195
(3) Instance based strategy: In this strategy, we also construct a document for each entity, but the words are from the instances related to that entity. For a class entity, words in the label, comment and property value of all its instances are extracted as the entity’s document; for a property entity, all the values it occurs in instances are extracted as the entity’s document. Then the similarity between two entities is calculated as in Metadata based strategy.
When combining the results of different matching strategies, we use a different method from which we used in OAEI 2008 and 2009. Instead of aggregating similarity values before extracting final alignment, we first extract alignment based on each individual strategy by threshold filtering method, and then combine alignments of different strategies together. A similarity propagation procedure based on structure information is performed to find more mappings. The similarity propagation procedure is implemented in iteration; in each iteration, the similarity is propagated from already found mappings to the rest candidate mappings, candidate mappings which get high similarity are then added to found mappings; this process is repeated until no more mapping is found. This combination method can generate alignments with very high precision with acceptable recall. Data Interlinking track
The DI (Data Interlinking) track is designed to test the ontology matching systems’ ability on link generation of LinkedData. There are five datasets, i.e. DailyMed, Diseasome, DrugBank, Sider and LinkedMDB, requested to be matched to related datasets in the LinkedData respectively. These data sets are all comes from the real world data and in relatively larger scale than the generated dataset. We choose four datasets in the domain of medicine to test our algorithm while exclude the linkedMDB dataset. According to our observations on the instance data, we split the information in the instance into six categories: the URL, the Meta Information, the Name, the string type information, the non-string type information and the neighboring information. Among the six categories the Name, which usually comes from the rdfs:label property or other ontology specific property such as foaf:name) is the most distinguishing feature to identify an instance. In addition, the natural language information and the neighboring instances are very useful, too. Thus we propose a vector based method for the DI track. We build two vectors referred to as Name Vector and Virtual Document for each instance. The Name Vector is constructed by accumulating the terms in the Name property values and setting the occurrence of each term as its weight. For Virtual Document, we first collect the terms of the each instance’s descriptions and annotations then fetch the local information of its neighboring instances to construct a comprehensive vector. Because the Virtual Document Space is much larger, we compute the tf-idf value of each term as its weight. The similarity between two instances is calculated as the weighted sum of their similarity (Cosine Distance) on two kinds of vectors respectively. However, this method is infeasible on large scale input because pair-wise comparisons on instances are too costly. Thus we introduce a candidate selection process. Only the instance pairs which are selected as candidate mappings are compared. Generally we use two rules for candidate selection: 1) instances with common terms in their Name Vectors; 2) instances with common top weighted terms in their Virtual Documents. To utilize the functionality, we build inverted index of instances for terms in Name Vector and top weighted terms in the Virtual Document. Consequently our algorithm
196
can generate the candidates very quickly and eliminate the meaningless comparisons between unrelated instances. Several experiment results show that the candidate selection will not eliminate the possible alignments in most of the cases. In the following phase of the algorithm, we may use the Meta Information and non-string type values as restrictions to filter the results according to the instance characteristics. For example, a common one is that those instances whose classes are not matched will be filtered out. At last a threshold is used on similarity for the final result. Totally speaking, this method is a generic and efficient method for instance matching. IIMB and PR track
Traditionally, information of individuals in an ontology is frequently utilized in supporting of schema matching. Inversely, information of schema is of equal importance in alignment of individuals that are sharing the same ontology structure. Thus, for the Instance Matching Track of this year, we take more about schema information, especially classes and properties, into consideration in aligning individuals.
For Instance Matching, our main idea is that we classify individuals by their classes, complete information of each individual as complete as possible, run matching algorithm for each class respectively, and compute similarity of two candidates based on weight-mean of properties assigned with specified weights. And the algorithm can be generalized as four consecutive phase:
Preprocessing: Read and store the schema information for further use. Build a local schema that connects properties and classes and implement it by learning information of individuals.
Information Complementation: Modify the information of each individual, aiming at making them as complete as possible. We defined some rules for judging the validity of values, as well as for solving the transformations in value, structure and logical. Reclassify individuals by recognizing and comparing properties they carry with those in classes, based on our local schema implemented in the previous phase.
Matching: Given the facts that different properties of individuals play quite different roles, and that every individual has its unique characteristic(s), for each property, we assign it with a specified weight and combine this weight with string-based similarity value computed under Edit Distance or Vector based algorithm. We assign the weight-mean of properties as the final similarity value.
Spread Similarity: In order to fully utilize the connection of individuals, we apply a similarity-flooding-like algorithm to spread the similarity.
1.3 Adaptations made for the evaluation
In order to deal with large scale data sets, we use an inverted index technique to accelerate the speed of locating and reading data.
1.4 Link to the system and parameters file
The RiMOM System can be found at http://keg.cs.tsinghua.edu.cn/project/RiMOM/
197
1.5 Link to the set of provided alignments (in align format)
The results of RiMOM for OAEI 2010 Campaign are available at http://keg.cs.tsinghua.edu.cn/project/RiMOM/OAEI2010/
2 Results
As introduced above, RiMOM participates in three tracks in OAEI 2010; we present the results and related analysis below.
2.1 Benchmark
There are 111 alignment tasks in benchmark data set; we divide these tasks into three groups: 1xx, 2xx, and 3xx. We compare the results of RiMOM in OAEI 2010 and OAEI 2009 [5] in Table 1. It can be observed that the performance of RiMOM in 1xx task continues to be perfect as last year; as for the 2xx task, the result of this year is better than that of last year, with regard to both precision and recall; the precision of 3xx increases this year, but the recall decreases, while the F1-measure is almost the same as last year. Overall, the precision, recall and F1-measure for the entire benchmark data set of RiMOM this year achieve 99% precision, 84% recall and an F1-measure of 91%. Compared with last year’s result, there are 6% improvement on precision, 2% improvement on recall and 4% improvement on F1-measure.
Table 1. Benchmark test results of RiMOM in OAEI 2010 and OAEI 2009 (Values are real precision and recall and not an average of precision and recall)
Test OAEI 2010 OAEI 2009 Prec. Rec. F1 Prec. Rec. F1
1xx 1.00 1.00 1.00 1.00 1.00 1.00 2xx 0.99 0.83 0.91 0.93 0.81 0.87 3xx 0.91 0.74 0.82 0.81 0.82 0.81 H-mean 0.99 0.84 0.91 0.93 0.82 0.87
2.2 DI track of IM@OAEI2010
We generate results for four of five datasets in the track except the LinkedMDB dataset. Since we are requested to mapping each dataset to several related datasets in LinkedData and these datasets not provided in the track, we download these datasets and transfer them into RDF format using Jena. As a result we cannot get some datasets such as STITCH because there is only a SPARQL endpoint for it. We also found there are many duplicate entries in the reference alignment of Sider and the namespace for DBpedia in the reference alignment of Drugbank is not uniform, we adjust these reference files to get the final result of our algorithm. We set the parameter of our algorithm as NameWeight = 0.6 and threshold = 0.55. The result of
198
Sider dataset is shown in Table 2. From the result we can see that according to the different characteristics of the instance file, the results may be very different: some are high in precision and some are high in recall. For those high in recall but low in precision, more careful filter may be added to the algorithm by studying the data. On the other hand, for those low in recall, the threshold may be cut down.
Table 2. The result of Sider Dataset DataSet DBpedia DailyMed Diseasome Drugbank TCM STITCH TOTAL
Precision 0.717 0.567 0.315 0.961 0.778 / 0.617 Recall 0.482 0.706 0.837 0.342 0.812 / 0.467
F-Measure 0.576 0.629 0.458 0.504 0.795 / 0.532
The result of DailyMed dataset is shown in Table 3. The result of our algorithm is extremely bad in the LinkedCT dataset. It generates a lot of results (up to 100,000) so that the precision is very low. Because of the dominance of LinkedCT results in the reference, our result in total is not good, too. According to our observation on the reference alignment of LinkedCT, they are automatically generated from the owl:SeeAlso property in the file. After reviewing some of our results, we found that many of our results are reasonable but some of the references are not, we think the reference alignment is not very complete and sound. However, our algorithm cannot generate good results from DBpedia means we need much more improvement on it. The other two datasets with LinkedCT reference, Diseasome and DrugBank are similar in results.
Table 3. The result of DailyMed dataset DataSet DBpedia LinkedCT TCM Sider TOTAL
Precision 0.246 0.070 0.159 0.567 0.085 Recall 0.293 0.235 0.535 0.706 0.296
F-Measure 0.267 0.107 0.123 0.629 0.132
Table 4. Results of IIMB IIMB_SMALL IIMB_LARGE
Dataset Prec. Rec. F1 Dataset Prec. Rec. F1 001 - 020 0.975 0.975 0.975 001 - 020 0.997 0.994 0.995 021 - 030 0.861 0.710 0.778 021 - 030 0.798 0.696 0.744 031 - 060 0.913 0.953 0.933 031 - 040 0.843 0.766 0.803 061 - 070 0.809 0.639 0.714 041 - 060 0.877 0.976 0.924 071 - 080 0.792 0.500 0.613 061 - 070 0.663 0.586 0.622
071 - 080 0.575 0.557 0.566
199
2.3 IIMB track of IM@OAEI2010
The result for IIMB_SMALL and IIMB_LARGE is shown in Table 4. As the number of datasets increases, the text-based information the dataset contains decrease while complex combination of modifications increase, thus the performance of our algorithm decreases since it is anyway fundamentally based on string comparison. We can also see that with the amount of instances grows, the influences brought by the noise increase, which do nothing but harm to effect of our algorithm.
2.4 PR track of IM@OAEI2010
PR track consists of three subtasks; the results for these tasks are shown in Table 5. It can be observed that RiMOM gets perfect performance on the first task; for the second task, RiMOM gets really good recall and the precision is 95.2%; for the last task, the precision and recall both decrease compared to the former two tasks.
Table 5. Results of PR Dataset Precision Recall F-Measure
Person11 - Person12 1.0 1.0 1.000 Person11 - Person12 0.952 0.99 0.971
Restaurant1 - Restaurant2 0.86 0.768 0.811
2.5 VLCR track
The purpose of VLCR task is to match three resources to each other, namely, the Thesaurus of the Netherlands Institute for Sound and Vision (called GTAA), the New York Times subject headings and DBpedia. Each resource consists of lots of instances: 142,000 in GTAA, 12,000 in NYT and 7,500,000 in DBpedia. Table 6 lists the number of the mapping we found.
Table 6. Result for VLCR task Dataset Number of mappings
NYT-DBpedia 9257 GTAA-DBpedia 68337
NYT-GTAA Direct mapping 4324 Indirect mapping 4487
Due to the lack of information, sometimes it is very difficult to match two instances in NYT and GTAA directly. Since we have mapped the two relatively small instance sets to DBpedia, it is possible to use the map results to get more maps between the two small one. Instances in NYT and GTAA matches to the same instance in DBpedia will be added to the final results.As shown in the table, NYT–DBpedia, GTAA-DBpedia and NYT-GTAA are three subtasks of VLCR task. Indirect matching find 163(rise by 3.7% ) new mappings in NYT-GTAA task.
200
3 General comments
By far instance matching, especially matching on real world instance is still a very challenging problem. Instance Matching is of great importance for bringing the ontology matching into practical use with its wide range of application scenarios. Instance matching shows its special characteristics compared with the conventional schema matching and the large scale nature of instance matching is a big obstacle to employ the existing methods. A relatively generic and efficient method for instance matching is in great need. The IMEI track of OAEI 2010 provides a good platform to test the instance matching algorithms and this area will attract more attention in the community.
4 Conclusion
In this paper, we present the results of RiMOM in OAEI 2010 Campaign. We participate in three tracks this year, including Benchmark, IMEI, and VLCR. We have presented the architecture of RiMOM system and described specific techniques used in this campaign. In this campaign, we design a new strategy combination method for benchmark tracks, and get better performance than last year. We particularly focus on the instance matching task; propose some new strategies for these tasks. The results illustrates that our system RiMOM can achieve good performance in both schema matching and instance matching tracks.
Acknowledgement: The work is supported by the National Natural Science Foundation of China (No.
60973102), the National Basic Research Program of China (973 Program) (No. 2007CB310803), the National High-tech R&D Program (No. 2009AA01Z138), it is also supported by IBM SUR joint project.
References
1. http://linkeddata.org/. 2. J. Li, J. Tang, Y. Li, and Q. Luo. RiMOM: A dynamic multi-strategy ontology alignment
framework. IEEE Transaction on Knowledge and Data Engineering, 21(8):1218–1232, Aug 2009.
3. Q. Zhong, H. Li, J. Li, G. Xie, and J. Tang. A Gauss Function based approach for unbalanced ontology matching. In Proc. of the 2009 ACM SIGMOD international conference on Management of data (SIGMOD’2009), Jul 2009.
4. F. Shi, J. Li, and J. Tang. Actively learning ontology matching via user interaction. In Proc. of the 8th International Conference of Semantic Web (ISWC’2009), Oct 2009.
5. X. Zhang, Q. Zhong, J. Li, J. Tang, G. Xie, and H. Li. RiMOM results for OAEI 2008. In Proc. of the Third International Workshop on Ontology Matching (OM’08), 2008.
201
Alignment Results of SOBOM for OAEI 2010
Peigang Xu, Yadong Wang, Liang Cheng, Tianyi Zang
School of Computer Science and Technology
Harbin Institute of Technology, Harbin, China
[email protected], [email protected], [email protected],[email protected]
Abstract. In this paper we give a brief explanation of how Sub-Ontology based
Ontology Matching (SOBOM) method gets the alignment results at OAEI2010.
SOBOM deal with an ontology from two different views: an ontology with is-a
hierarchical structure O’ and an ontology with other relationships O’’. Firstly,
from the O’ view, SOBOM starts with a set of anchors provided by a linguistic
matcher. And then it extracts sub-ontologies based on the anchors and ranks
these sub-ontologies according to their depths. Secondly, SOBOM utilizes
Semantic Inductive Similarity Flooding algorithm to compute the similarity of
concepts between different sub-ontologies derived from the two ontologies
according the depth of sub-ontologies to get concept alignments. Finally, from
the O’’ view, SOBOM gets relationship alignments by using the concept
alignment results in O’’. The experiment results show SOBOM can find more
alignment results than other compared relevant methods.
1 System presentation
Currently more and more ontologies are distributedly built and used by different
organizations. And these ontologies are usually light-weighted [1] containing lots of
concepts especially in biomedicine, such as anatomy taxonomy NCI Thesaurus. The
Sub-ontology based Ontology Matching (SOBOM) is designed for matching light-
weight ontologies that has is-a hierarchy as their backbones. It matches an ontology
202
from two views: O’ and O’’ that are depicted in Fig. 1. The unique feature of our
method is combining sub-ontology extraction with ontology matching.
1.1 State, purpose, general statement
SOBOM is developed to match ontology automatically for general purpose. Based on
two different views, we design three elementary matchers in current version. The first
one is a anchor generator which is used to find anchors; the second one is a structure
matcher SISF (Semantic Inductive Similarity Flooding) which is inspired by Anchor-
Prompt [3] and SF [4] algorithms and is exploited to flood similarity among concepts.
The last one is a relationship matcher which utilizes the results of SISF to get
relationship alignments. In addition, a Sub-ontology Extractor (SoE) is integrated into
SOBOM to extract sub-ontologies according to the anchors got by linguistic matcher
and rank them by their depths descendingly. Overall SOBOM is a sequential method,
so it does not care how to combine the results of different matchers. The overview of
the method is illustrated in Fig. 2.
O 'O ''O
Fig. 1. Two views of an ontology in SOBOM
'O 'O ''O
2_OSub
2_OSub
1_OSub
1_OSub
Fig. 2. The processing overview of SOBOM
For simplicity, we define some notations used in the report.
203
Ontology: An ontology O consists of a set of concepts C, properties (relations) R,
instances I, and axioms A. We use entity e to denote either Cc or Rr . Each
relation r has a domain and range defined as following:
}|{)( riprelationshthehavingandCccrDomain ii
}|{)( rofvaluebecanandCccrRange ii
Anchor: An anchor is defined as a pair of assumed equivalent non-leaf concepts
across ontologies. Given two ontologies 1O , 2O , 11 Oc , 22 Oc , if 21 cc (means that c1 is identical with c2),and c1, c2 are both not leaf nodes in the hierarchies of O1
and O2 respectively, then an anchor, X is defined as a pair of concepts 21,cc .Sub-Ontology: Let ontology O = (C,R, HC,HR,). A sub-ontology Ox is a subset of
O whose elements all come from O, called Ox = (C1,H1C), where CC HHCC 11 , ,
x is the root of HC. Indeed, a sub-ontology in our method is a hierarchical taxonomy,
and its root is an anchor concept.
Sub-ontology Depth. The depth of sub-ontology Ox is the maximal length of path
from the anchor x to the root ri of the taxonomy HiC which contains it in the original
ontology O.
OHHrrxMaxODepth Ci
Ciii
x ,,...,)(
1.2 Specific techniques used
SOBOM aims to provide high quality of 1:1 alignments between concept and
property pairs. We have implemented SOBOM algorithm in java and integrated three
distinguishing constitutional matchers. They are independent components in core
matcher library of SOBOM. Due to the space limitation, we only describe the key
features of them. The details can be found in the related paper [8].
Our anchor generator is based on the local context of an entity in ontology. In
details, the local context of an entity including the following aspects: the
textual information (label, id, comments and so on), the structure information
(the number of super or sub concepts, the number of constraints) and the
individual information (the number of individual if existing). We consider
204
that the local context of an entity can express the meaning of it. Consequently,
we get three similarity matrixes respectively, and we choose the maximal of
them as the final results.
SISF uses the RDF statement to represent the ontology and utilizes the
anchors to inducting the construction of similarity propagation graph for the
sub-ontologies. SISF handles the ontology from the view O’ and only
generate concept-concept alignment.
R-matcher is a relationship matcher base on the definition of the ontology. It
combines the linguistic and semantic information of a relation. From the O’’
view, it utilizes the is-a hierarchy to extend the domain and range of a
relationship and uses the result of SISF to generate the alignment between
relationships.
More importantly, SoE is integrated into SOBOM and extracts sub-ontologies
according to the anchors [5,6]. SoE ranks extracted sub-ontologies according to their
depths. As we extract sub-ontologies for ontology matching, the rules of extracting
sub-ontology in SoE are as following: only sub-concepts of anchor are included in the
sub-ontology. In other words, a sub-ontology is a taxonomy which has anchor as root.
If one of the two concepts in an anchor is a leaf node in the original ontology, we
do not use SISF to deal with it actually. Because this phenomenon just represents a
one-to-many mapping. After extracting sub-ontologies, SOBOM will match these sub-ontologies according to their depth in original ontology. We first match the sub-ontologies with larger depth value. By using SoE, SOBOM can reduce the scale of
ontology and make it easy to operate sub-ontologies in SISF.
1.3 Adaptations made for the evaluation
We don’t make any specific adaptation for the tests in the OAEI 2010 campaign. All
the alignments outputted by SOBOM are based on the same set of parameters.
205
1.4 Link to the system and parameters file
The current version of SOBOM is available at:
http://mlg.hit.edu.cn:8080/Ontology/Download.jsp, and the parameters setting is
illustrated in the reading me file.
1.5 Link to the set of provided alignments (in align format)
We deploy our matcher as a web service, our web service name is:
eu.sealsproject.omt.ws.matcher.AlignmentWSImpl. The endpoint of our web service
can be found at: http://mlg.hit.edu.cn:8080/SOBOMService/SOBOMMatcher?wsdl.
2 Results
In this section, we describe the results of SOBOM algorithm against the benchmark,
directory and anatomy ontologies provided by the OAEI 2010 campaign. We use
Jena-API to parse the RDF and OWL files. The experiments were carried out on a PC
running Windows vista ultimate with Core 2 Duo processors and 4-gigabyte memory.
2.1 Benchmark
On the basis of the nature, we can divide the benchmark dataset into five groups:
#101-104, #201-210, #221-247, #248-266 and #301-304. SOBOM is a sequential
matcher. If the linguistic matcher gets no results, SOBOM will produce no result. We
described the performance of our SOBOM algorithm over each group and overall
performance on the benchmark test set in Table 1.
#101-104 SOBOM plays well for these test cases.
#201-210 In this group, some linguistic features of candidate ontologies are
discarded or modified, their structures are quite similar. SOBOM is a sequential
matcher, our anchor generator matches concepts based on their local context not only
206
the linguistic information. So, although without linguistic information SOBOM also
gets relatively high precision and recall.
#221-247 The structures of the candidate ontologies are altered in these tests.
However, SOBOM discovers most of the alignments from the linguistic perspective
via our anchor generator, and both the precision and recall are pretty good.
#248-266 Both the linguistic and structural characteristics of the candidate
ontologies are changed heavily, so the tests in this group might be the most difficult
ones in all the benchmark tests. So, SOBOM does not very well.
#301-304 This test group are four real-life ontologies of bibliographic
references. SOBOM can only find equivalence alignment relations.
Table 1. The performance on the benchmark
101-104 201-210 221-247 248-266 301-304
Precision 1.0 0.99 0.99 0.87 0.77
Recall 1.0 0.85 0.99 0.57 0.70
Compared to our previous results (OAEI2009), the recall of every group is highly
improved. This is enhanced by our redesigned anchor generator.
2.2 Anatomy
The anatomy real world test bed covers the domain of body anatomy and consist of
two ontologies, Adult Mouse Anatomy (2247 classes) and NCI Thesaurus (3304
classes). These are relatively large compared to benchmark ontologies. This type
ontologies is what SOBOM suitable for handling, it generated 268 sub-ontologies and
1249 alignments between MA and NCI, consumed 19min3s to complete the matching
task. It is obvious that most of the alignment appears in the leaf nodes in ontologies
(834 leaf node alignments). The experiment result shows in Table 2.
Table 2. The performance of SOBOM on the anatomy test
Leaf node
alignments
Sub-
ontologies
Total
Alignments
Time
consuming
207
NCI 834 268 1249 19min3s
MA
2.3 Conference
There are 120 pairs of ontologies in this track. Most of them are blind tests (i.e.
there no reference alignment available). The whole results are available at:
http://seals.inrialpes.fr/platform/;jsessionid=FD1E3A5CE8DA43C1D52DB21079EA
ECF3?wicket:bookmarkablePage=:eu.sealsproject.omt.ui.Results&endpoint=http://21
9.217.238.162:8080/SOBOMService/SOBOMMatcher?wsdl&evaluationID=http://21
9.217.238.162:8080/SOBOMService/SOBOMMatcher?wsdl2010/10/03+02:09:03&tr
ack=Conference+Testsuite.
3 General comments
In this section, we want to introduce comments on the results of SOBOM algorithm
and the way to improve it.
3.1 Comments on the results
Strengths SOBOM deals with ontology from two different views and combines
results of every step in a sequential way. If the ontologies have regular literals and
hierarchical structures, SOBOM can achieve satisfactory alignments. And it can avoid
missing alignment in many partitioning matching methods as illustrated in [7].
Weaknesses SOBOM needs anchors to extract sub-ontologies. So it depends on the
precision of anchors. In current version, we use a linguistic matcher to get anchor
concept, if the literals of concept missed, SOBOM will get bad results.
208
3.2 Discussions on the way to improve the proposed system
SOBOM can be viewed as a frame of ontology matching. So many independent
matchers can be integrated into it. Now, we have enhanced the anchor generator by
not considering the textual information but also the structure information. Our next
plan is to integrate a more powerful matcher to produce anchor concepts or develop a
new method to get anchor concepts. Meanwhile, we plan to develop a mapping
debugging method to refine the results of SOBOM.
4 Conclusion
Ontology matching is very important part of establishing interoperability among
semantic applications. This paper reports our participation in OAEI2010 campaign.
We present the alignment process of SOBOM and describe the specific techniques for
ontology matching. We also show the performance in different alignment tasks. The
strengths and the weaknesses of our proposed approach are summarized and the
possible improvement will be made for the system in the future. We propose a brand
new algorithm to match ontologies. The unique feature of our method is combining
sub-ontology extraction with ontology matching based on two different views of an
ontology.
References
1. Fausto Giunchiglia and Ilya Aihrayeu : Lightweight Ontologies. Technical Report. (2007)
DIT-07-071.
2. G.Stoilos, G. Stamou, S. Kollias: A string metric for ontology alignment, In Proc. Of the 4th
International Semantic Web Conference(ISWC’05). (2005) 623-637.
3. N.F. Noy, M.A. Musen: Anchor-PROMPT: using non-local context for semantic matching,
In Proc. Of IJCAI2001 Workshop on Ontology and Information Sharing, ( 2001) 63-70.
209
4. S. Melnik, H.G. Molina and E. Rahm: Similarity Flooding: A Versatile Graph Matching
Algorithm, In Proc 18th Int’l Conf. Data Eng. (ECDE’02) (2002) 117-128.
5. Julian Seidenberg and Alan Rector: Web Ontology Segmentation: Analysis, Classification
and Use, WWW2006, (2006).
6. H. Stuckenschmidt and M. Klein: Structure-Based Partitioning of Large Class Hierarchies. In
Proc of the 3rd International Semantic Web Conference (2004).
7. W. Hu, Y. Qu: Block matching for ontologies, In Proc of the 5th International Semantic Web
Conference, LNCS, vol. 4273, Springer (2006) 300-313.
8. P.G. Xu, H.J. Tao: SOBOM: Sub-ontology based Ontology Matching. To be appear.
210
TaxoMap alignment and refinement modules: Resultsfor OAEI 2010
Faycal Hamdi1, Brigitte Safar1, Nobal B. Niraula2, and Chantal Reynaud1
1 LRI CNRS UMR 8623, Universite Paris-Sud 11, Bat. G, INRIA Saclay
2-4 rue Jacques Monod, F-91893 Orsay, France
[email protected] The University of Memphis
Memphis, TN, USA
Abstract. TaxoMap is an alignment tool which aims to discover rich correspon-
dences between concepts (equivalence relations (isEq), subsumption relations
(isA) and their inverse (isMoreGnl) or proximity relations (isClose)). It performs
an oriented alignment (from a source to a target ontology) and takes into ac-
count labels and sub-class descriptions. This new implementation of TaxoMapuses a pattern-based approach implemented in the TaxoMap Framework helping
an engineer to refine mappings to take into account specific conventions used in
ontologies.
1 Introduction
TaxoMap was designed to retrieve useful alignments for information integration be-
tween different sources. The alignment process is then oriented from ontologies that
describe external resources (named source ontology) to the ontology (named target on-
tology) of a web portal. The target ontology is supposed to be well-structured whereas
source ontology can be a flat list of concepts.
TaxoMap makes the assumption that most semantic resources are based essentially
on classification structures. This assumption is confirmed by large scale ontologies
which contain rich lexical information and hierarchical specification without describing
specific properties or instances. Then, to find mappings we use the following available
elements: labels of concepts and hierarchical structures.
The new implementation of TaxoMap introduces a step of refinement of mappings
(the alignment results) which extends the alignment process and completes it.
We take part to three tests. We hope the new step of refinement helps us to perform
better in terms of precision of generated mappings.
2 Presentation of the System
Our system is composed of two elements: TaxoMap, the alignment tool and TaxoMapFramework, an environment allowing to specify and perform refinement treatments ap-
plied on the prior obtained mappings.
211
2.1 State, Purpose and General Statement
TaxoMap has been designed to align owl ontologies O = (C, H). C is a set of concepts
characterized by a set of labels and H is a subsumption hierarchy which contains a set
of isA relationships between nodes corresponding to concepts. The alignment process is
an oriented process which tries to connect the concepts of a source ontology OS to the
concepts of a target ontology OT . The correspondences found are equivalence relations
(isEq), subsumption relations (isA) and their inverse (isMoreGnl) or proximity relations
(isClose).
To identify these correspondences, TaxoMap implements techniques which exploit
the labels of the concepts and the subsumption links that connect the concepts in the
hierarchy. The morpho-syntactic analysis tool, TreeTagger [1], is used to classify the
words of the labels of the concepts and to divide them into two classes, full words and
complementary words, according to their category and their position in the labels. At
first the repartition between full and complementary words is used by a similarity mea-
sure that compares the tri-grams of the labels of the concepts [2] and gives more weight
to the common full words. Then it is also used by the alignment techniques. For exam-
ple, one technique named LabelInclusion generates an isA mapping between cs and
ctmax if (1) the concept ctmax is the concept of OT having the highest similarity value
with the concept cs of OS , (2) one of the labels of ctmax is included in one of the labels
of cs, (3) all the words of the included label of ctmax are classified as full words by
TreeTagger.
Given a concept cS of the ontology source OS , our similarity measure identify the
concept ctmax of the target ontology OT which have the highest similarity with cS .
The alignment techniques are then used to decide if the concept cS can be effectively
aligned with this concept ctmax and which relation should be established between the
two concepts, or whether, another concept of OT must be chosen. A proposed mapping
belongs to a single method, a concept of OS can be aligned at most with one concept
of OT . In contrast, the concepts of OT may be involved in several proposed alignments.
The main methods used to extract mappings between a concept cs in OS and a
concept ct in OT are:
– Label equivalence: An equivalence relationship, isEq, is generated if the similar-
ity between one label of ctmax and one label of cs is greater than a threshold
(Equiv.threshold).
– Label inclusion (and its inverse): If one of the labels of ctmax is included in one
of the labels of cs, and if all words of included label are full words, we propose a
subclass relationships < cs isA ctmax >. Inversely, if one of the labels of cs is in-
cluded in one of the labels of ctmax, we propose the relationships < cs isMoreGnlctmax >.
– High lexical similarity: If the similarity measure of ctmax is greater than a thresh-
old (HighSim.threshold) and if one of its labels shares at least two full words in
common with one of the labels of cs, without being including in the labels of cs,
the heuristic generates the relationship < cs isClose ctmax >.
212
– Reasoning on similarity values: Let ctmax and ct2 be the two concepts in OT with
the highest similarity measure with cs, the relative similarity is the ratio of ct2
similarity on similarity ctmax. If the relative similarity is lower than a threshold
(isA.threshold), one of the two following techniques can be used:
• the relationship < cs isClose ctmax > is generated if the similarity of ctmax
is greater than a threshold (isClose.thresholdMax).
• an isA relationship is generated between cs and the father of ctmax if the sim-
ilarity of ctmax is greater than a second threshold (isA.thresholdMax).
– Best similarity: If none of the above techniques is applicable, the relationship < cs
isClose ctmax > is generated if the similarity of ctmax is greater than a threshold
(Better.thresholdMax).
– Property similarity: Two classes cs and ct are likely to be aligned if they share the
same properties. Our property similarity is computed using Degree of Commonality
Coefficient presented in [8].
Mappings identified by TaxoMap are generated in the Alignment format used as
a standard in the OAEI campaign. We added to this format the information about the
names of the techniques that generated mappings. The aim is to facilitate the specifi-
cation of treatments exploiting the mappings generated by those techniques. All these
pieces of information are stored in a relational mappings database which can then be
queried using SQL queries. This allows, in particular, to present the generated mappings
to the expert in the validation phase, technique by technique.
In the OAEI campaigns, only equivalence relations are evaluated in the alignment
contest. This has important implications on our results:
1. None of the mappings generated by the label inclusion techniques that lead to a
subsumption relation isA is considered as such. Most of them are wrong if they are
converted into equivalence relation.
2. Moreover, while it is natural to consider that several different concepts can be relied
by a subsumption relation to the same concept, the conversion of subsumption rela-
tion into equivalence relation leads to the creation of multiple equivalence relations
for a same concept.
3. As a concept of OT must have only one equivalent concept in OS in OAEI cam-
paigns, if we consider the mappings leading to subsumption or proximity relations,
all mappings which connect a concept of OS to a concept of OT which is already
involved in an equivalence relation are false.
We will see in the next section how the TaxoMap refinement module [7] will allow
us to remove these incorrect mappings.
2.2 TaxoMap Framework
We proposed an environment allowing to specify and perform treatments applied on the
prior obtained mappings. At first, this environment will be used to improve the quality
of an alignment provided by TaxoMap. Subsequently, it will be used for other treatments
based on mappings as enriching, restructuring or merging ontologies.
213
An important feature of the approach is to allow a declarative specification of treat-
ments based on particular alignment results, concerning particular ontologies and using
a predefined vocabulary. Treatments which can be specified depend on the character-
istics of the concerned ontologies and on the task to be performed (at first mapping
refinement and subsequently ontology merging, restructuring, enriching). These treat-
ments are thus associated to independent specification modules, one for each task, each
having their own vocabulary. The approach is extensible and a priori applicable to any
treatment based on alignment results.
We present the Mapping Refinement Pattern Language (MRPL) used to specify
mapping refinement pattern. This language differs from the one defined in [3] espe-
cially because it includes patterns which test the existence of mappings generated by
alignment techniques.
The vocabulary of MRPL contains:
– a set of predicate constants. We distinguish three categories of predicate constants:
the predicate constants relating to the type of techniques applied in the identification
of a mapping by TaxoMap, the predicate constants expressing structural relations
between concepts of a same ontology, the predicate constants expressing termino-
logical relations between labels of concepts.
– a set of individual constants: {a, b, c, ...}– a set of variables: {x, y, z, ..., } where is an unnamed variable used to represent
parameters which do not need to be precised.
– a set of built-in predicates: {Add Mapping, Delete Mapping}– a set of logical symbols: {∃, ∧, ¬}
MRPL allows the definition of a context part which must be satisfied to make the
execution of a pattern possible, and of a solution part which expresses the process to
achieve when the context part is satisfied.
Context part of pattern
The context part tests (1) the technique used to identify the considered mapping,
(2) the structural constraints on mapped elements, for example, the fact that they are
related by a subsumption relation to concepts verifying or not some properties, or (3)
the terminological constraints, for example, the fact that the labels of a concept are in-
cluded in the labels of other concepts. These conditions are represented using formulae
built from predicate symbols. So, we distinguish three kinds of formula according to
the kind of predicate symbols used.
The formulae related to the type of techniques applied in the identificationof a mapping by TaxoMap. By testing the existence in the mappings database of
a particular relation generated by a given technique, we build formulae that implic-
itly test the conditions for the application of this technique. For example the formula
isAStrictInclusion(x, y) tests the existence of a mapping isA generated between two
214
concepts x and y using the technique searching LabelInclusion, t2. It validates im-
plicitly at the same time all the conditions for the application of t2, i.e. (1) the concept
y is the concept of OT having the highest similarity value with the concept x of OS ,
(2) one of the labels of y is included in one of the labels of x, and (3) all the words
of the labels of y are classified as full words by TreeTagger. TaxoMap includes several
alignment techniques. Thus, several predicate symbols leading to formulae of that kind
are needed. More formally, let:
RM = {isEq, isA, isMoreGnl, isClose}, the set of correspondence relations
used by TaxoMap,
T = {t1, t2, .........}, the set of techniques.
TM , the table storing generated mappings in the form of 4-tuple (x, y, r, t) where
x ∈ CS , y ∈ CT , r ∈ RM , t ∈ T . The pairs of variables (x, y) which can instantiate
these formulae will take their values in the set (x, y) | (x, y, r, t) ∈ TM . The predicate
symbols necessary for the task of refinement presented in this paper are isEquivalentand isAStrictInclusion the semantics of which are the following:
– isEquivalent(x, y) is true iff ∃(x, y, isEq, t1) ∈ TM
– isAStrictInclusion(x, y) is true iff ∃(x, y, isA, t2) ∈ TM
– mapping(x, y) is true iff ∃(x, y, , ) ∈ TM
The formulae expressing structural relations between concepts x and y of thesame ontology O = (C, H). Since the aim of TaxoMap is the alignment of taxonomies,
the structural relations considered here are subsumption relations. If the approach was
used with another alignment tool, other relations could be considered. Note that the
instances of variables in these formulae will be constrained, either directly because
they instantiate the previous formulae, related to the type of the applied techniques, or
indirectly by having to be in relation with other instances.
– isSubClassOf(x, y, O) is true ⇔ isA(x, y) ∈ H– isParentOf(x, y,O) is true ⇔ isA(y, x) ∈ H– conceptsDifferent(x, y) is true ⇔ ID(x) �= ID(y) with ID(x) is the identifier
of the concept x.
The formulae expressing terminological relations between the labels of the con-cepts: not detailed here because not used in the examples of this paper.
Solution part of pattern
A context part is associated to a solution part which is a set of actions to be per-
formed. This set of actions is modeled by a conjunction of built-in predicates executed
in a database. The built-in predicates are defined as follows:
– Add Mapping(x, y, r) has the effect of adding a tuple to the table TM which be-
comes TM ∪ {(x, y, r, t)} where r and t are fixed in the treatment condition by
instantiating the predicate corresponding to the type of technique associated with
the considered mapping.– Delete Mapping(x, y, ) has the effect of removing a tuple from the table TM
which becomes TM − {(x, y, , )}.
215
Mapping Refinement Pattern used in OAEI
Pattern-1: This pattern concerns mappings generated by the technique t1, connect-
ing by an equivalence relationship a concept y of the target ontology OT with a concept
x of the source ontology OS . Because a concept y of the target ontology OT must be
involved in at most one equivalence relation, mappings involving y and obtained from
other techniques than t1 should be removed.
Context part of Pattern-1:∃x∃y (isEquivalent(x, y)∧ ∃z (mapping(z, y) ∧ conceptDifferent(z, x) ))
Solution part of Pattern-1:Delete Mapping(z, y, )
Fig. 1. Illustration of Pattern-1
Pattern-2: For the anatomy subtask 4, if we know a set of reference mappings, and
we know that the pair (x, y) belongs to this set, we could express a new refinement
pattern to remove generated mappings that relie to the concept y of the target ontology,
concepts z of the source ontology other than x.
We should define the new predicate referenceMapping(x, y) as follow:
referenceMapping(x, y) is true iff ∃(x, y) ∈ Set Reference Mapping.
Context part of Pattern-2:∃x∃y (referenceMapping(x, y)∧ ∃z (mapping(z, y) ∧ conceptDifferent(z, x) ))
Solution part of Pattern-2:Delete Mapping(z, y, )
2.3 Link to the system and parameters file
TaxoMap requires:
– Java (Version 1.6 and above )3
The version of TaxoMap used in 2010 contest can be downloaded from:
– http://www.lri.fr/ hamdi/TaxoMap/TaxoMap.html
3 http://java.sun.com
216
2.4 Link to the Set of Provided Alignments
The alignments produced by TaxoMap are available at the following URLs:
http://www.lri.fr/˜hamdi/OAEI10/
3 Results
3.1 Benchmark Tests
Since our algorithm only provides mapping for concepts, the recall is low even for the
reference alignment. The overall results show a slight improvement over those the last
year.
3.2 Anatomy Test
The anatomy real world case is to match the Adult Mouse Anatomy (denoted by Mouse)
and the NCI Thesaurus describing the human anatomy (tagged as Human). Mouse has
2,744 classes, while Human has 3,304 classes. We considered Human as the target on-
tology as is it well structured and larger than Mouse. TaxoMap performs the alignment
in about 12 minutes.
Table 1. Results of TaxoMap in the different tasks
Task#1 Task#2 Task#3
# Computed Mappings 1 449 1 127 2 357
Precision without Refinement 0.779 0.929 0.477
# Removed Mappings 226 32 945
# Submitted Mappings 1 223 1 095 1 412
Precision 0.924 0.956 0.838
Δ Precision with Refinement + 0.145 + 0.027 + 0.361
Recall 0.743 0.689 0.774
F-measure 0.824 0.801 0.802
As only equivalence relationships will be evaluated in the alignment contest, we did
not use this year the techniques which generate isA relationship (except in the subtask
3) and we change isClose mapping to equivalence. In addition, we use the refinement
pattern described above to delete mappings between a concept of the target ontology
that was already aligned with an equivalence mapping. As a result, we found fewer
mappings than last year but the precision is better [4].
3.3 Directory Test
The directory task consists of Web sites directories like Google, Yahoo! or Looksmart.
Two modalities are proposed this year:
217
– Small tasks: includes 4,639 tests represented by pairs of OWL ontologies.– Single task: includes only one matching task. The source and the target ontologies
to be matched contain 2854 and 6555 concepts respectively.
TaxoMap takes about 40 minutes to complete each modality.
4 General Comments
4.1 Results
The new version of TaxoMap improves significantly the results on the previous ver-
sion of TaxoMap in terms of runtime and precision of generated mappings. The new
implementation offers extensibility and modularity of code. TaxoMap can be parame-
terized by the language used in ontologies, the choice of used techniques and different
thresholds.
4.2 Future Improvements
The following improvements can be made to obtain better results:
– To take into account all concepts properties instead of only the hierarchical ones.– To use WordNet as a dictionary of synonymy. The synsets can enrich the termino-
logical alignment process if an a priori disambiguation is made.– To develop the remaining structural techniques which proved to be efficient in last
experiments [5] [6].
5 Conclusion
This paper reports our participation to OAEI campaign with the new implementation
of TaxoMap. Our participation in the campaign allows us to test the robustness of Tax-oMap, the new implemented techniques and the efficacity of the new refinement module
in TaxoMap Framework.
References[1] Schmid H. Probabilistic Part-of-Speech Tagging Using Decision Trees, International Confer-
ence on New Methods in Language Processing (1994)[2] Lin, D. An Information-Theoretic Definition of Similarity. ICML. Madison. (1998) 296–304[3] Scharffe, F.: Correspondence Patterns Representations. PhD thesis, Univ. of Innsbruck, 2009.[4] Hamdi, F., Safar, B., Niraula, NB. and Reynaud, C. TaxoMap in the OAEI 2009 alignment
contest, Proceedings of the ISWC’09 Workshop on Ontology Matching OM-09 (2009)[5] Reynaud, C. and Safar, B. When usual structural alignment techniques don’t apply, The
ISWC’06 Workshop on Ontology matching (OM-06), (2006)[6] Reynaud, C. and Safar, B. Exploiting WordNet as Background Knowledge,The ISWC’07
Workshop on Ontology Matching (OM-07), (2007)[7] Hamdi, F., Reynaud, C. and Safar, B., Pattern-based Mapping Refinement, 17th International
Conference on Knowledge Engineering and Knowledge Management, EKAW (2010)[8] Reul, Q. and Jeff Z. Pan, KOSIMap: Ontology alignments results for OAEI 2009, Proceed-
ings of the ISWC’09 Workshop on Ontology Matching OM-09 (2009)
218
Towards a UMLS-based silver standardfor matching biomedical ontologies
E. Jimenez-Ruiz1, B. Cuenca Grau2, I. Horrocks2, and R. Berlanga1
1 Universitat Jaume I, Spain, {ejimenez,berlanga}@uji.es2 University of Oxford, UK, {berg,ian.horrocks}@comlab.ox.ac.uk
Abstract. We propose a silver standard based on the UMLS Metathe-saurus to align NCI, FMA and SNOMED CT. This silver standard aimsat being exploited within the OAEI and SEALS Campaigns.
1 Motivation
The UMLS Metathesaurus (UMLS-Meta) [1] is currently the most comprehen-sive effort for integrating independently-developed medical thesauri and ontolo-gies. UMLS-Meta is being used in many applications, including PubMed andClinicalTrials.gov. The integration of new UMLS-Meta sources combines auto-matic techniques, expert assessment, and auditing protocols (see [2] for a reviewof current methods). In its 2009AA version, UMLS-Meta integrates more thanone hundred thesauri and ontologies, including SNOMED CT, FMA and NCI,and contains more than 6 million entities. UMLS-Meta provides a list with morethan two million unique identifiers (CUIs). Each CUI can be associated to en-tities belonging to different sources. Pairs of entities from different sources withthe same CUI are synonyms and hence can be represented as an equivalencemapping. Thus, UMLS-Meta mappings could be considered as a silver standardto align ontologies such as SNOMED CT, FMA or NCI. The Ontology Align-ment Evaluation Initiative (OAEI) could be benefited from UMLS-Meta so thatit could be used as the input dataset for a new challenging track within the eval-uation campaign. However, in our previous work [3, 4], we showed that UMLS-Meta contains a significant number of logic errors when the rich semantics of theontology sources is taken into account together with the UMLS-Meta mappings.
2 Method and Results
Our experiments were based on the UMLS-Meta version 2009AA and the corre-sponding versions of FMA (version 2.0), NCI (version 08.05d) and SNOMED CT(version 20090131), which contain 66,724, 78,989 and 304,802 entities, respec-tively. After extracting the relevant parts of UMLS, we obtained 3,024 mappingaxioms between FMA and NCI, 9,072 between FMA and SNOMED CT and19,622 between SNOMED CT and NCI. Note that mappings are considered asOWL 2 axioms (see [5, 4]).
219
When reasoning over each of the source ontologies independently, all theirentities were found satisfiable. However, after the respective integrations viaUMLS-Meta mappings, we obtained a huge number of unsatisfiable entities,namely 5,015 when integrating FMA and NCI, 16,764 with FMA and SNOMEDCT, and 76,025 with SNOMED CT and NCI.
We designed three logic-based principles [3, 4], namely conservativity prin-ciple, consistency principle and locality principle, to automatically detect andrepair conflictive set of mappings. After the assessment, our automatic meth-ods removed 570 (19%) of the mappings between FMA and NCI, 4,077 (45%)of those between FMA and SNOMED CT and 13,358 (63%) of those betweenSNOMED CT and NCI. When reasoning with the new revised mapping sets, wefound only 2 unsatisfiable entities when integrating FMA and NCI, 44 for FMAand SNOMED CT, and none for SNOMED CT and NCI. These remaining errorswere analyzed with our semi-automatic tool ContentMap [5] and they requiredto repair inherent incompatibilities between the source ontologies (see [4]).
3 Discussion
UMLS-Meta represents a reference of correspondences among ontologies, how-ever, the direct integration of these ontologies with UMLS-Meta mappings leadsto a huge number of unintended logical consequences. We propose instead a re-vised set of UMLS-Meta mappings which could be considered as a silver standardto evaluate ontology matching techniques over FMA, NCI and SNOMED CT.The proposed silver standard and related sources are available for download inhttp://krono.act.uji.es/people/Ernesto/umlsassessment.
The silver standard could be improved by ontology matching tools if theyare able to find new valid correspondence which do not lead to logic errors. Wealso intend to design less aggressive techniques in order to preserve the maxi-mum number of UMLS-Meta mappings. Additionally, ontology sources could berevised before the automatic assessment in order to smooth the incompatibilites.
References
[1] Bodenreider, O.: The unified medical language system (UMLS): integratingbiomedical terminology. Nucleic acids research 32(Database issue) (2004) 267–270
[2] Geller, J., Perl, Y., Halper, M., Cornet, R.: Special issue on auditing of terminolo-gies. Journal of Biomedical Informatics 42(3) (2009) 407–411
[3] Jimenez-Ruiz, E., Cuenca Grau, B., Horrocks, I., Berlanga, R.: Towards a logic-based assessment of the compatibility of UMLS sources. In: Proceedings of theSWAT4LS workshop. Volume 559 of CEUR Workshop Proceedings. (2009)
[4] Jimenez-Ruiz, E., Cuenca Grau, B., Horrocks, I., Berlanga, R.: Logic-based assess-ment of the compatibility of UMLS ontology sources. Accepted for publication inBMC Journal of Biomedical Semantics (2010)
[5] Jimenez-Ruiz, E., Cuenca Grau, B., Horrocks, I., Berlanga, R.: Ontology integra-tion using mappings: Towards getting the right logical consequences. In: Proc. ofESWC. Volume 5554 of LNCS., Springer-Verlag (2009) 173–187
220
From French EHR to NCI ontology via UMLSPaolo Besana1, Marc Cuggia1, Oussama Zekri2, Annabel Bourde1,Anita
Burgun1
1 Université de Rennes 1, 2 Centre Régional de Lutte Contre Le Cancer - EugèneMarquis
Abstract. Clinical trials are required for evaluating new therapies anddiagnostic techniques. We developed a system based on OWL and SWRLaimed at suggesting the clinical trials to which a patient could be en-rolled, evaluating on real clinical trials and patients from the UniversityHospital of Rennes. This paper presents the method employed to map theFrench expressions from the patient data to terms in the NCI ontology.
1 IntroductionClinical trials are fundamental for testing new therapies or diagnostic techniques.Patients are enrolled to a clinical trial if they match its eligibility criteria. Wedeveloped a work aimed at suggesting the clinical trials to which a patient couldbe enrolled [2]. We represent patient data from the Electronic Health Record(EHR) and the eligibility criteria using the NCI Thesaurus [4], an oncology-specific ontology developed by the National Institutes of Health. The projectevaluation is based on the clinical trials and patients’ data from the CentreHospitalier Universitaire of Rennes (France) . This paper presents the work doneto automatically map the French text of some of the EHR fields to terms fromthe NCI ontology.
2 Problem descriptionIn order to evaluate the system we extracted 4 real clinical trials active during2009 in the University Hospital of Rennes, and we selected 486 patients that wereassessed for the trials during the same year. Both clinical trials and data were inFrench. The rules have been converted manually, while the size of the patientsdata required an automatic processing for at least some of the fields. This paperpresents the method employed to automatically map the French expressions ofsome of the EHR fields to terms from the NCI ontology.
3 MethodA patient’s record contains 44 relevant fields. We focussed on the fields containingsingle expressions. In particular, we chose the field specifying the site of thetumor, very relevant for the recruitment. We extracted the possible differentsource values Vsrc =
{v1src, ..., v
nsrc
}of the field . All patients are admitted in
urology, and all the cancers are related to urology: the possible values for thefield are only 28. Of these values, 16 are composed by more than a single word(for example: ’col de l’uterus’ meaning cervix). In particular 5 are specification ofa position within an organ. We compared three methods for translation: MESH(Medical Subject Headings, terminology used to index PubMed publications)in French1, Google translate and Wikipedia. In order to find the NCI concept,1 http://terminologiecismef.chu-rouen.fr/
221
Method Translation found Correct translation CUIs found NCI found NCI correct
MESH 16 - 57% 16- 57% 13 - 46% 12 - 42% 12 - 42%
Wikipedia 20 - 71% 20 - 71% 21 - 75% 18 - 64% 15 - 53%
Google 28 - 100% 26 - 92% 14 - 50% 9 - 32% 7 - 25%Table 1. Mapping results for the three different methods on 28 French terms. A trans-lation is found if an expression is returned. The expressions are evaluated manually.The CUIs are concepts in UMLS found by string matching, and the NCI terms arelinked to CUIs. The correctness of the NCI terms is checked manually.
we used the UMLS meta-thesaurus (Unified Medical Language System) [3]. InUMLS each concept is identified by a unique key (CUI). A concept is linked tomany terms from different terminologies, including NCI. Each term in UMLS isdefined by a semantic type (such as Disease or Body Part). Table 1 shows theresults for the different methods, discussed more in detail below.
Using Google Translate Google translate was queried for each of the terms inVsrc. The only error in translation is for a term with homonyms in differentdomains. The translations were used to query the UMLS by string. To improverelevance, we filtered the results by semantic type. The terms specifying a po-sition in an organ could not be matched directly to defined terms in any of theterminology contained in UMLS. The found CUIs were then filtered to extractthe terms from NCI.
Using MESH French Each term in Vsrc is queried on the MESH French website.The result contains the corresponding English term from the original version ofMESH. The English term is then used to query UMLS (filtering by string andby MESH terminology). Of the 16 found translations, 3 were terms belonging todifferent semantic types and therefore discarded. The resulting CUIs are queriedto find the corresponding NCI term.
Using Wikipedia Inspired by the work in [1], each source term is queried inwikipedia French. A possible set of pages is returned. The pages are selectedusing their wikipedia categories. The filtered pages are used to extract the cor-responding English page title. The mapping to NCI follows the same mechanismexplained for Google translate. The use of categories to filter unrelated pagesimproves precision.
References
1. O. Collin B. Gaillard, M. Boualem. Query translation using wikipedia resourcesfor analysis and disambiguation. In Proceeding of EAMT 2010, Conference of the
European Association for Machine Translation, 2010.2. P. Besana, M. Cuggia, O. Zekri, A. Bourde, and A. Burgun. Using semantic web
technologies for clinical trial recruitment. In International Semantic Web Conference
2010, volume 6414. Springer, 2010.3. O. Bodenreider. The unified medical language system (UMLS): integrating biomed-
ical terminology. Nucleic Acids Research, 32(Database Issue):D267, 2004.4. N.F. Noy, S. de Coronado, H. Solbrig, G. Fragoso, F.W. Hartel, and M.A. Musen.
Representing the NCI Thesaurus in OWL DL: Modeling tools help modeling lan-guages. Applied ontology, 3(3):173–190, 2008.
222
From Mappings to Modules: Using Mappings to IdentifyDomain-Specific Modules in Large Ontologies
Amir Ghazvinian, Natalya F. Noy, Mark A. Musen
Stanford University, Stanford, CA 94305, US
{amirg,noy,musen}@stanford.edu
Using Mappings to Identify Modules. Ontology modularization is an active area
of research in the Semantic Web community [3]. With the emergence and wider use
of very large ontologies, in particular in fields such as biomedicine, more and more
developers need to extract meaningful modules of these ontologies to use in their appli-
cations. Researchers have also noted that many ontology-maintenance tasks would be
simplified if we could extract modules from ontologies. These tasks include ontology
matching: If we can separate ontologies into modules, we can simplify and improve on-
tology matching. We study a complementary problem: Can we use existing mappings
between ontologies to facilitate modularization?
Methods. Figure 1 illustrates our method: First, we generate mappings from a
source ontology to the target ontology that we wish to modularize. Next, we cluster
mappings within the target ontology. Finally, we use mapping clusters to identify mod-
ules within an ontology.
Validation and Analysis. We validate and analyze our approach by applying our
methods to identify modules for NCI Thesaurus [1] and SNOMED-CT [2], two popu-
lar and large biomedical ontologies. As domain-specific ontologies for modularization,
we used 141 ontologies in BioPortal.1 Our process extracted 71 modules for NCI The-
saurus and 68 modules for SNOMED-CT. We examined modules and their representa-
tive terms in order to understand the types of modules that our algorithm creates and to
determine whether or not these modules are likely to be useful in an application setting.
Figure 2 shows an example, a module of NCI Thesaurus that is relevant to electrocar-
diograms (EKG) using the Electrocardiography Ontology. The module consists of 61
classes, representative samples of which are shown in the figure. Of the 61 classes in
this module, 41 (67%) are mapping targets.
Discussion and Conclusions. Our approach uses mappings between ontologies in
order to extract domain-specific modules from large ontologies based on their mappings
to smaller ontologies. In our experiments with NCI Thesaurus and SNOMED-CT, using
the ontologies from BioPortal as the sources for mappings, we have identified a number
of useful modules. We found that one of the key hurdles that we must overcome, is to
find a way to determine how good a particular module is. Indeed, the same problem
is true for most modularization approaches [3]: many authors discuss computational
properties of their modules, but do not evaluate how useful these modules are to users. In
our case, the requirements for extraction are driven by domain coverage of the module
rather than by its computational or structural properties. Thus, the problem of evaluating
whether the module satisfies the user requirements is similar to the problem of ontology
1 http://bioportal.bioontology.org
223
evaluation in general: how do we know that an ontology is useful for a specific class
of applications? We plan to submit the modules that we have identified to BioPortal
to enable the user community to use the modules in their applications, review them
and comment on them. However, our initial evidence, which we present in this paper,
indicates that our approach can indeed find interesting domain-specific modules.
Fig. 1. The process of identifying modules using mappings between ontologies. A: mappings
between a source ontology and a modularization target. B: two clusters returned by clustering
the mappings. One cluster is light gray in color while the other is dark gray. When determining a
module based on these clusters, we discard the light gray cluster since the mapping targets within
that cluster are too sparse. C: the process of pruning the ontology subtree for the remaining
cluster, which we use to create the module. We begin at each leaf and traverse the tree toward
the root, removing all classes that are not mapping targets or direct children of mapping targets.
Once we reach such a class, we stop pruning along that branch. D: the final module.
Fig. 2. A portion of the module
that we identified within NCI
Thesaurus that represents the
domain of the Electrocardiog-
raphy Ontology. The classes in
gray represent mapping targets
and classes in white represent
classes that were not mapping
targets, but are included in the
module through our algorithm.
References
1. N. Sioutos, S. de Coronado, M. Haber, F. Hartel, W. Shaiu, and L. Wright. NCI Thesaurus:
A semantic model integrating cancer-related clinical and molecular information. Journal ofBiomedical Informatics, 40(1):30–43, 2007.
2. K. Spackman, editor. SMOMED RT: Systematized Nomenclature of Medicine, Reference Ter-minology. College of American Pathologists, Northfield, IL, 2000.
3. H. Stuckenschmidt, C. Parent, and S. Spaccapietra, editors. Modular Ontologies: Concepts,Theories and Techniques for Knowledge Modularization. Springer-Verlag, Berlin, 2009.
224
225
226
Concept abduction for semantic matchmaking indistributed and modular ontologies
Viet-Hoang VU, Nhan LE-THANH
KEWI - Laboratory I3S - CNRSNice Sophia Antipolis University, France.
Abstract. Recently, attempts have been done to formalize SemanticMatchmaking, the process of finding potential matches between demandsand supplies based on their logical relations with reference to a commonontology, in Description Logics (DLs). We extend the formalization toPackages-based Description Logics (P-DLs), modular extensions of DLs,in order to perform the matchmaking operation in contexts where de-scriptions of demands and supplies are specified in different terminolo-gies.
Keywords: semantic matchmaking, package-based description logics,distributed ontology, abductive reasoning, tableaux-based methods.
1 Introduction
Recently, several attempts have been done to formalize Semantic Matchmaking,the process of finding potential matches based on logical relations with referenceto a common ontology, in Description Logics (DLs)[4]. Demands and supplies arethus represented as concepts in order to use different inference mechanisms toevaluate possible matches between them according to the semantic relationships.We extend this formalization to the context of Packages-based Description Logics(P-DLs)[2] to allow descriptions of demands and supplies can be specified indifferent, distributed but interconnected ontologies. For performing the semanticmatchmaking in such settings, we develop a distributed reasoning algorithmbased on tableau calculus to compute concept abduction, an abductive reasoningservice developed specifically for this kind of operation[4], in P-DLs.
2 Distributed abductive reasoning in Package-basedDescription Logics
Description Logics (DLs)[1] are a family of logic-based languages for represent-ing and reasoning about the knowledge of a domain. Package-based DescriptionLogics (P-DLs) are extensions of DLs to represent distributed and modular on-tologies. In P-DLs, a knowledge base (KB) is thus considered as a collection ofcomponents called packages. In each package, along with its local terms, the us-age of foreign terms imported from other packages are permitted, allowing localknowledge of packages can be reused elsewhere.
227
Concept Abduction is a novel non-monotonic inference task proposed fordescription logics to evaluate potential matches between demands and supplies.Given two concepts C, D and a TBox T such that C � D ��T ⊥, this reasoningservice allows to find a concept H (hypothesis) such that C � H ��T ⊥ andC � H �T D .
Extending Concept Abduction to the distributed context of P-DLs, in placeof a single TBox T , we have a set of packages Σ = {Pi}. Let Pw ∈ Σ besome witness package, the computing of H now need to be done with respectto Pw . To do that, we devise a distributed algorithm which is based on thefederated reasoning technique developed for P-DLs[2] and the uniform tableaux-based method[3]. The procedure consists of two stages:
1. At first, we try to build multiple, federated local tableaux for C �w D1. If allsuch tableaux contain obvious contradiction, then either C is unsatisfiableor the subsumption holds w.r.t Σ as witnessed by Pw and thus no abductionis needed.
2. On the contrary, for all tableaux which are consistent, we compute conceptexpressions that, when added to the tableaux, will eventually generate somecontradiction. If these contradictions imply in fact the subsumption betweentwo concepts C and D, they will be parts of the finding hypothesis H.
3 Conclusions
We have developed a distributed tableaux-based algorithm for solving the con-cept abduction problem in the distributed and modular context of Package-basedDescription Logics. This allows to perform the Semantic Matchmaking in situ-ations where demands and supplies are specified with reference to different on-tologies. In the future, we would like to optimize this method to achieve betterperformance.
References
1. Franz Baader, Diego Calvanese, Deborah L. McGuinness, Daniele Nardi, and Pe-ter F. Patel-Schneider, editors. The Description Logic Handbook: Theory, Imple-mentation, and Applications. Cambridge University Press, 2003.
2. Jie Bao, Doina Caragea, and Vasant Honavar. A tableau-based federated reasoningalgorithm for modular ontologies. In In IEEE/WIC/ACM International Conferenceon Web Intelligence, pages 404–410. IEEE Press, 2006.
3. S. Colucci, T. Di Noia, E. Di Sciascio, F. M. Donini, and M. Mongiello. A uniformtableaux-based approach to concept abduction and contraction in aln. In Contrac-tion in ALN . In Proc. of DL-04 (2004), CEUR Workshop Proceedings, page 104,2004.
4. Tommaso Di Noia, Eugenio Di Sciascio, and Francesco M. Donini. Semantic match-making as non-monotonic reasoning: A description logic approach. Journal of Ar-tificial Intelligence Research, 29:307, 2007.
1 ’�w’ is used to denote that the inference is performed from the local point of viewof Pw.
228
LingNet: Networking Linguistic and Terminological Ontologies
Wim Peters Department of Computer Science
University of Sheffield, UK [email protected]
Abstract. Linguistic description and linguistically-based re-engineering tech-niques depend on standardized models. LingNet aims at the alignment of mul-tiple standard description models from the terminological, linguistic and locali-zation fields.
1 Motivation
Linguistic and terminological standards are in daily use for the purpose of linguistic and terminological resource creation (term banks, dictionaries, translation memories etc.). Since there is a variety of standard models available. Examples of terminological and linguistic models are the ISO standard initiatives ISO16620 and the Lexical Markup Framework (LMF) [1]. Furthermore, the purpose of e.g. LexInfo [2] and the Linguistic Information Repository (LIR) [3] is to associate multilingual linguistic knowledge with conceptual ontology elements. In the translation memory area standards such as TMX1 (Translation Memory eXchange) and XLIFF2 (XML Localization Interchange File Format) are widely used.
In order to enable interdisciplinary re-use and complementarity it is necessary to establish interoperability between their vocabularies in a principled way. LingNet3 is a model for mapping linguistic and terminological (standard) information in a distributed fashion. The LingNet model adopts a number of modelling decisions from the literature: a knowledge-based, formalism-independent metamodel for capturing semantic alignments between ontologies for linguistic/terminological description [4], references to external mapping patterns for structural mappings [5], and a the integration of a lexicalization relation for linking linguistic/terminological resource to ontological concepts [6]. The novelty of the LingNet model lies in its combination of selected mapping me-thods and its application to the linguistic/terminological domain. The basic structure of LingNet is illustrated in figure 1 below.
1 http://www.lisa.org/tmx/ 2 http://docs.oasis-open.org/xliff/v1.2/os/xliff-core.pdf 3 first version: http://www.gate.ac.uk/ns/ontologies/LingNet/LingNet-v0.1.owl.
229
Fig. 1. The LingNet Model
References
1. Francopoulo, G., Monte George, Nicoletta Calzolari, Monica Monachini, Nuria Bel, Mandy Pet, Claudia Soria: LMF for multilingual, specialized lexicons. In: LREC, Genova, Italy (2006)
2. Buitelaar, P., Cimiano, P., Haase, P. and Sintek, M.: Towards Linguistically Grounded Ontologies, in: The Semantic Web: Research and Applications, 6th European Semantic Web Conference, proceedings of ESWC 2009, Heraklion, Crete, Greece, Springer Berlin / Heidelberg, Volume 5554 (2009)
3. Peters, W. Montiel-Ponsoda, E. Aguado de Cea, G.: Localizing Ontologies in OWL. In: Proceedings of the ISWC07 OntoLex workshop, Busan, Korea (2007)
4. Brockmans, S., Haase, P., Stuckenschmidt, H.: Formalism-Independent Specification of Ontology Mappings - A Metamodeling Approach, In: Robert Meersman, R., Tari, Z. et al. (eds), OTM 2006 Conferences, Springer Verlag, Montpellier, France (2006)
5 Scharffe, F. Euzenat, J. and Fensel, D.: Towards design patterns for ontology alignment. In R.L. Wainwright and H. Haddad (eds.): Proceedings of the 2008 ACM Symposium on Applied Computing (SAC), Fortaleza, Ceara, Brazil, March 2008: 2321-2325 (2008)
6 Picca, D., Gangemi, A., and Gliozzo, A.: LMM: an OWL Metamodel to Represent Heterogeneous Lexical Knowledge. In Proc. of the International Conference on Language Resources and Evaluation (LREC), Marrakech, Morocco (2008)
230
Aggregation of Similarity Measures in OntologyMatching
Lihua Zhao1 and Ryutaro Ichise2
1National Institute of Informatics, Tokyo, Japan, [email protected]
2National Institute of Informatics, Tokyo, Japan, [email protected]
Abstract. This paper presents an aggregation approach of similarity
measures for ontology matching called n-Harmony. The n-Harmony mea-
sure identifies top-n highest values in each similarity matrix to assign a
weight to the corresponding similarity measure for aggregation. We can
also exclude noisy similarity measures that have a low weight and the
n-Harmony outperforms previous methods in our experimental tests.
1 Introduction
Ontology matching is a promising research field that discovers similarities be-tween two ontologies and is widely used in applications such as semantic web,biomedical informatics and software engineering. Most of current ontology match-ing systems combine different similarity measures. For instance, the authors in[4] applied the Ordered Weighted Average(OWA) to combine similarity measuresand Ichise[3] proposed a machine learning approach to aggregate 40 similaritymeasures.
Harmony[5] measure is a state-of-the-art adaptive aggregation method thatassigns a higher weight to reliable and important similarity measure and a lowerweight to those fail to map similar ontologies. The harmony weight for a simi-larity measure is calculated according to the number of the highest values in thecorresponding similarity matrix. However, the harmony measure has drawbackswhen there exist other similarity measures that are as important as the oneswith the highest similarity value. Hence, we extended the harmony measure byconsidering top-n values in each row and column of similarity matrices and wecall this method as n-Harmony measure. The top-n is calculated according tothe number of concepts in two ontologies. Our extended n-Harmony considersmore values in similarity matrices and only aggregates similarity measures thathave a high harmony weight.
2 n-Harmony Measure
We applied 13 different similarity measures for aggregation which include 4string-based, 1 structure-based and 8 WordNet-based similarity measures[2]. The
final aggregated similarity matrix is∑
k(nHk×SMatrixk(Os,Ot))
|SMatrix| , where nHk is n-
Harmony weight and SMatrix is the similarity matrix of each similarity measure
231
between ontology Os and Ot. Before combining the similarity matrices, we re-move min(L-1, nH × L) lowest values in each row and column of similaritymatrix, where L is the minimum number of concepts in two ontologies and nHrepresents harmony weight of corresponding similarity matrix. Furthermore, onlythose similarity matrices with a high harmony weight are aggregated for the finalsimilarity matrix. The final decision of whether a ontology pair is matching ornot depends on the final similarity matrix and manually tuned threshold.
Directory data sets3 and Benchmark data sets4 from OAEI5 are tested withour system. The n-Harmony measure returns best result on Directory data setswhen the threshold is 0.45 with 0.86 recall and 0.70 F-measure while the origi-nal harmony measure returns 0.81 recall and 0.68 F-measure. This result is alsobetter than the results of best systems in OM2009[1], such as ASMOV whichreaches 0.65 recall and 0.63 F-measure on the Directory data sets. On the Bench-mark data sets, n-Harmony performs the same as the original harmony measureor returns slightly better recalls and F-measures than harmony. Comparing withthe ASMOV, n-Harmony performs almost the same on data sets #101-104 and#221-247 and returns higher precisions on #302-304, but slightly lower recalls.
3 Conclusions and Future Work
Experimental results show that our n-Harmony outperforms original harmonymeasure on most of the Directory and Benchmark data sets and also comparablewith the best systems attended in OM2009. However there are still rooms to im-prove our n-Harmony measure by exploring advanced structure-based similaritymeasures and by investigating automatic threshold selection method rather thanmanually tuning the threshold to find out the best performance.
References
1. Euzenat, J., et al.: Results of the ontology alignment evaluation initiative 2009. In:
Proceedings of the 4th International Workshop on Ontology Matching. pp. 82–135
(2009)
2. Euzenat, J., Shvaiko, P.: Ontology matching. Springer-Verlag, Heidelberg (2007)
3. Ichise, R.: Machine learning approach for ontology matching using multiple concept
similarity measures. In: Proceedings of the 7th IEEE/ACIS International Conference
on Computer and Information Science. pp. 340–346 (2008)
4. Ji, Q., Haase, P., Qi, G.: Combination of similarity measures in ontology matching
using the owa operator. In: Proceedings of the 12th International Conference on In-
formation Processing and Management of Uncertainty in Knowledge-Based Sytems.
pp. 243–250 (2008)
5. Mao, M., Peng, Y., Spring, M.: An adaptive ontology mapping approach with neural
network based constraint satisfaction. In: Web Semantics: Science, Services and
Agents on the World Wide Web. pp. 14–25. No. 8 (2010)
3http://oaei.ontologymatching.org/2009/directory/
4http://oaei.ontologymatching.org/2009/benchmarks/
5http://oaei.ontologymatching.org/
232
Using Concept and Structure Similarities forOntology Integration
Xiulei Liu12, Payam Barnaghi1, Klaus Moessner1, and Jianxin Liao21Centre for Communication Systems Research University of Surrey
Guildford, Surrey, GU2 7XH, United Kingdom2State Key Laboratory of Networking and Switching Technology
Beijing University of Posts and Telecommunications Beijing 100876, P.R. of China{xiulei.liu,p.barnaghi,k.moessner}@surrey.ac.uk
Abstract. We propose a method to align different ontologies in sim-ilar domains and then define correspondence between concepts in twodifferent ontologies using the SKOS model.
Introduction. Recently ontologies are created to provide knowledge represen-tation. They use common representation languages such as OWL, but there aremany heterogeneous ontologies [1–3]. In this paper we first propose a lexicaland structural analysis and compute the concept similarity as a combinationof attributes, second use the SKOS model to define correspondence betweenconcepts[4].Ontology Alignment Framework.To perform the matching between conceptsin different ontologies, we focus both on syntactical and text in entity descrip-tions and also their semantic structure in the ontology representations. Thisprocess, illustrated in the block diagram shown in Figure 1, is divided into twomain sub-tasks: Alignment and SKOS translation. The inputs are two ontologiesand result of the process is an SKOS-based ontology that contains automaticallydefined associations.The alignment task analyses lexical and structural attributesof ontologies to automatically produce associations between concepts. The re-lation is defined: R(A,B) =< A,B,Relation,S(A,B) > where A and B areontology concepts, Relation describe semantic relations between these conceptswhich have five types: equal beIncluded, include, disjoint, related, and S(A,B) issimilarity measure for two concepts based on their structure and lexical analysis.
Fig. 1. The ontology alignment process
233
2
Fig. 2. Snapshot of the specified properties in the integrated ontology
Defining SKOS-based Associations. After identifying possible relations be-tween concepts, they are imported based on the SKOS model. This will providean interconnection between two ontologies based on standard set of properties de-fined in the SKOS model. The SKOS mapping properties include skos:closeMatch,skos:exactMatch, skos:broadMatch, skos:narrowMatch and skos:relatedMatch. Theproperties maintain a mapping between SKOS concepts adapted from schemes.The relations in concept pairs defined in the previous section are based on synsetrelations in WordNet. They are obtained according to accessing the extendedsynset collection for each representative word that describes entities and calcu-lating structural similarity We will map between synset and SKOS relations. Byapplying these mappings, the final product of the ontology integration processwill include assertion axioms in which the related concepts from different ontolo-gies are linked to each other based on SKOS relations. The integrated ontologywill be a collection of concepts and properties from both ontologies and willalso include the SKOS association properties. Figure 2 illustrates a part of theSKOS relations and concept alignment between two ontologies from the dataset(a complete set of our evaluation results using OAEI2008 dataset can be accessedfrom: http://tinyurl.com/38veolh).Acknowledgement.The work is partially supported by the m:Ciudad projectfunded by the European 7th Framework Programme, contract number: 215007.Xiulei Liu’s and Jianxin Liao’s research is supported by Chinese National ScienceFund for Distinguished Young Scholars (No. 60525110) and Chinese National 973Program (No. 2007CB307100, 2007CB307103).
References
1. T. L. Bach, J. D. Bo, and R. Lara, “Knowledgeweb consortium, state of the art onontology (d2.2.3),” 2004.
2. P.Shvaiko and J.Euzenat, “A survey of schema-based matching approaches,” in J.Data Semantics IV, 2005, pp. 146–171.
3. Y. Kalfoglou and M. Schorlemmer, “Ontology mapping: The state of the art,” 2003.4. A. Miles and S. Bechhofer, “(editors) skos simple knowledge organization
system reference, w3c recommendation 18 august 2009.” [Online]. Available:http://www.w3.org/TR/skos-reference/
234
Flexible Bootstrapping-Based Ontology Alignment
Prateek Jain, Pascal Hitzler, Amit P. Sheth
Kno.e.sis Center, Wright State University, Dayton, OH
Abstract. BLOOMS (Jain et al, ISWC2010, to appear) is an ontology alignment
system which, in its core, utilizes the Wikipedia category hierarchy for establish-
ing alignments. In this paper, we present a Plug-and-Play extension to BLOOMS,
which allows to flexibly replace or complement the use of Wikipedia by other on-
line or offline resources, including domain-specific ontologies or taxonomies. By
making use of automated translation services and of Wikipedia in languages other
than English, it makes it possible to apply BLOOMS to alignment tasks where
the input ontologies are written in different languages.
1 Introduction
In this work, we present the extension of BLOOMS called Plug-n-Play BLOOMS (PnP
BLOOMS). PnP BLOOMS allows users to plug in a datasource of choice to assist in
the task of ontology matching while utilizing the core BLOOMS approach. Thus, PnP
BLOOMS can be customized by plugging in other information sources best suited to the
needs of the domain and the user. At the same time it has been developed in a modular
fashion to allow for de-coupling of the main matching technique with the auxiliary
source. The system is available from http://wiki.knoesis.org/index.php/BLOOMS.
The paper is organized as follows. In Section 2, we present our proposed solution,
Section 3 contains a minimalistic evaluation which shows the feasibility of our approach
and finally we conclude in Section 4.
2 Our Solution
We present the BLOOMS approach following [1], modified as appropriate to cater for
our plug-n-play extension. BLOOMS constructs a forest (i.e., a set of trees) TC (which
we call the BLOOMS forest for C) for each matching candidate class name C – in the
original approach, this roughly corresponds to a selection of supercategories of the class
name in the sense of the Wikipedia clas shierarchy. Comparison of the forests TC and
TB for matching candidate classes C and B then yields a decision whether or not (and
with which of the candidate relations) C and B should be aligned.
We have implemented an extension of the BLOOMS framework from [1] to support
’plug-n-play’ methodology. Plug-n-Play methodology provides users the flexibility to
customize the framework to use the best auxiliary datasource based on their needs and
application scenario. The framework allows users to plug-in the auxiliary data source
which is best suited to the target domain. This is in part possible due to the nature of
the BLOOMS approach, which relies on the comparison of nodes in the categorization
tree.
235
Table 1. Comparison of BLOOMS and PnP BLOOMS on French ontologies of the Benchmark
track of OAEI 2009.
Ontology f-measure BLOOMS f-measure PnP BLOOMS
206 0.53 0.76
207 0.58 0.77
210 0.50 0.72
Avg. 0.54 0.75
3 Evaluation
In order to evaluate our plug-n-play approach for ontology matching, we utilized the
French language ontologies from the Benchmark track of OAEI 2010 initiative.1 We uti-
lized these ontologies since It allows us to compare the performance of PnP BLOOMS
with the old version of BLOOMS. Further, it allows us to check the plug-n-play version
of BLOOMS with respect to ease of use and functionality. Table 1 illustrates the com-
parative performance of PnP BLOOMS with BLOOMS. The table clearly illustrates
the advantage obtained by using the PnP approach. The effectiveness of the BLOOMS
PnP approach is demonstrated by an increment of 42% over the previous version of
BLOOMS.
4 Conclusion and Future Work
In this paper we have presented an extension of our recent work on BLOOMS [1] which
allows users to plug-in auxiliary sources of their choice as oracles in the task of ontology
matching. It differs from the previous version of BLOOMS which was restricted to
using English language Wikipedia to help in this task. The effectiveness of the approach
is demonstrated by the increment of 42% over the previous version of BLOOMS. The
flexible plug-n-play based approach allows users to customize the system based on their
needs and the domain and language of the ontologies. Using PnP BLOOMS, we are also
planning on investigating the use of ’cocktails’ of auxiliary data sources for the task of
ontology matching.
Acknowledgement This work is funded primarily by NSF Award:IIS-0842129, titled
”III-SGER: Spatio-Temporal-Thematic Queries of Semantic Web Data: a Study of Ex-
pressivity and Efficiency”. Pascal Hitzler acknowledges support by the Wright State
University Research Council.
References1. Jain, P., Hitzler, P., Sheth, A.P., Verma, K., Yeh, P.Z.: Ontology Alignment for Linked Open
Data. In: Proceedings of the 9th International Semantic Web Conference, ISWC 2010, Shang-
hai, China, November 7-11, 2010 (To Appear), Springer-Verlag (2010) 402–417 Available
from http://knoesis.org/library/resource.php?id=844.
1 http://oaei.ontologymatching.org/2009/
236
Semantic Matching of Ontologies
Christoph Quix, Marko Pascan, Pratanu Roy, David Kensche
Informatik 5 (Information Systems), RWTH Aachen University, Germany
http://www.dbis.rwth-aachen.de
Introduction The discovery of semantic relationships such as subsumption and dis-
jointness is still a challenge in ontology matching [6]. Existing methods use logical
reasoning over computed equivalence relationships or machine learning based on lex-
ical and structural features of ontology elements [7, 3]. While these methods deliver
good results for some cases, they are limited to the information contained in the input
ontologies to be matched. Therefore, background knowledge in form of an additional
ontology may be useful to detect semantic relationships. In existing approaches, the
identification of an appropriate ontology as background knowledge is often a task left
for the user. We present two enhanced approaches for identifying semantic relation-
ships. The first one is based on background knowledge; in contrast to other approaches,
it is able to identify a background ontology automatically. The second approach builds
on existing machine learning methods for identifying semantic relationships. First eval-
uation results for these methods and combined approaches show that the integration of
these methods is reasonable as more semantic relationships are identified.
Semantic Matching using Background Knowledge It has been shown in previous
works that using an ontology as background knowledge can improve the match result
[1]. The selection of the background ontology is obviously an important step in such an
approach. While earlier works either relied on the user to provide such an ontology [1],
or used very general upper ontologies (e.g. SUMO-OWL, [5]), our approach is able to
select the background ontology automatically. The idea is illustrated in fig. 1 and 2. For
the input ontologies S and T, we generate keyword queries for a web search engine (e.g.,
Google or Swoogle), and for our local ontology repository. The external search engine
is only used if the local repository does not contain an appropriate ontology. When a
background ontology O is found it can be used for matching. In addition to the direct
alignment Adir, two alignments AO,S and AO,T , between the input ontologies and
the background ontology, are computed. Then, for each pair of correspondences from
AO,S and AO,T , existence of a relationship (i.e., equivalence, subsumption) between
the model elements from O is determined. If that is the case, a new correspondence
between the concepts from S and T can be inferred. All the correspondences found in
this way are called semantic matches (Asem). Eventually, the final result is created by
building the union of Adir and Asem.
Using Machine Learning for Semantic Matching Our second, complementary method
uses machine learning to identify semantic relationships. We implemented an approach
similar to the method presented in [7]. The computation of subsumption relationships
is considered as a binary classification task, i.e., a concept pair is classified into two
possible classes: subsumption and not-subsumption relationship.
Because ontologies are usually hierarchical structures, the subsumption relation-
ships found in input ontologies can be used as training examples, making the process
of classifier training independent of alignment computation. Each training example is a
237
S TInput
OntologiesQSearch
Ontologies(from the web)
g
QS QT
KeywordQueries
gEngine
(from the web)
RepositoryOOOO
Fig. 1. Selection of the background ontology
S T
O
Adir
AO,S AO,T
Asem
Fig. 2. Matching using a background ontology
0 6
0,8
1,0 Initial
ML
0,0
0,2
0,4
0,6
0,8
1,0
cmt confOf cmt ekaw cmt iasted cmt sigkdd confOf ekaw
Initial
ML
BGK
Comb.1
Comb.20,0
0,2
0,4
0,6
0,8
1,0
cmt confOf cmt ekaw cmt iasted cmt sigkdd confOf ekaw
Initial
ML
BGK
Comb.1
Comb.2
Fig. 3. Recall for OAEI conference track
concept pair (ci, cj), where ci � cj and both concepts belong to a single ontology, i.e.,
the source or target ontology of a matching task. We use distinct words, extracted from
both source and target ontologies, as features for the machine learning method. In order
to represent the concept pairs in the feature space, each concept ci of a concept pair is
described by a set of feature space words that can be found in its neighborhood, consti-
tuting the concept’s context document Dci. The notion of the concept’s neighborhood
can be defined in various ways. In our implementation Dci is created from words found
in: name, label, comment, instances, data and object properties of ci, direct sub or super
concepts of ci, concepts in union or intersection definitions of ci and equivalent con-
cepts of ci. The context documents of concepts are translated to feature vectors which
are used as input for the machine learning method (currently, C4.5 decision tree). We
use an optimization to avoid the classification of all concept pairs.
Conclusion The presented approaches have been integrated into our generic matching
system GeRoMeSuite [4]. To evaluate our approach, we used semantic precision and
recall as measures [2] and data sets from the oriented track of OAEI 2009. We created
also two combined match configurations using both approaches. The results in fig. 3
show very good results for the combined approaches.
Acknowledgments: This work is supported by the Umbrella Cooperation Programme
(http://www.umbrella-coop.org/).
References
1. Z. Aleksovski, M. Klein, W. ten Kate, F. van Harmelen. Matching Unstructured Vocabularies
Using a Background Ontology. Proc. EKAW. 2006.2. J. Euzenat. Semantic precision and recall for ontology alignment evaluation. Proc. IJCAI, pp.
348–353. 2007.3. Y. R. Jean-Mary, E. P. Shironoshita, M. R. Kabuka. Ontology matching with semantic verifi-
cation. Journal of Web Semantics, 7(3):235–251, 2009.4. D. Kensche, C. Quix, X. Li, Y. Li. GeRoMeSuite: A System for Holistic Generic Model
Management. Proc. VLDB, pp. 1322–1325. 2007.5. V. Mascardi, A. Locoro, P. Rosso. Automatic Ontology Matching via Upper Ontologies: A
Systematic Evaluation. IEEE Trans. on Knowl. & Data Eng., 22:609–623, 2010.6. P. Shvaiko, J. Euzenat. Ten Challenges for Ontology Matching. Proc. ODBASE, pp. 1164–
1182. 2008.7. V. Spiliopoulos, G. A. Vouros, V. Karkaletsis. On the discovery of subsumption relations for
the alignment of ontologies. Web Semantics, 8(1):69–88, 2010.
238
Ontology Mapping Neural Network: An
Approach to Learning and Inferring
Correspondences among Ontologies
Yefei Peng1�, Paul Munro1, and Ming Mao2
1 University of Pittsburgh, Pittsburgh PA 15206, USA,[email protected], [email protected],
2 SAP Labs, Palo Alto CA 94304, USA,[email protected],
The Ontology Mapping Neural Network (OMNN) extends the ability of Iden-tical Elements Neural Network(IENN) and its variants’ [4, 1–3] to represent andmap complex relationships. The network can learn high-level features commonto different tasks, and use them to infer correspondence between the tasks. Thelearning dynamics of simultaneous (interlaced) training of similar tasks interactat the shared connections of the networks. The output of one network in responseto a stimulus to another network can be interpreted as an analogical mapping.In a similar fashion, the networks can be explicitly trained to map specific itemsin one domain to specific items in another domain. A more detailed version ispublished on the main conference [5].
The network architecture is shown in Figure 1. Ain and Bin are input sub-vectors for nodes from ontology A and ontology B respectively. They share onerepresentation layer ABr. RAin represents relationships from graph A; RBin
represents relationships from graph B. They share one representation layer Rr.In this network, each to-be-mapped node in graph is represented by a sin-
gle active unit in input layers (Ain, Bin) and output layers (Aout, Bout ). Forrelationships representation in input layer (RAin, RBin), each relationship isrepresented by a single active unit. The network shown in Figure 1 has multiplesub networks shown in the following list.
1. NetAAA : {Ain-ABr-XAB; RAin-RRA-XR }-H1-W -H2-VA-Aout;2. NetBBB : {Bin-ABr-XAB; RBin-RRB-XR }-H1-W -H2-VB-Bout;3. NetAAB : {Ain-ABr-XAB; RAin-RRA-XR }-H1-W -H2-VB-Bout;4. NetBBA : {Bin-ABr-XAB; RBin-RRB-XR }-H1-W -H2-VA-Aout;
Selected OAEI 3 benchmark tests are used to evaluate OMNN approach.Wilcox test is performed to compare OMNN with the other 12 systems par-ticipated in OAEI 2009 on precision, recall and f-measure. The result is shownin Figure 1. Green means OMNN is significantly better than the system; Redmeans the system is significantly better than OMNN. Yellow means no signifi-cant difference. Significance is defined as p − value < 0.05. It shows that OMNN
� The author is working at Google now. Email: [email protected] http://oaei.ontologymatching.org/
239
Fig. 1. Proposed network architecture and Results
has better F-measure than 9 of the 12 systems, OMNN’s recall is significantlybetter than 10 of the systems. It should be noted that p-value< 0.05 meansthere is significant difference between two systems compared, then detailed datais visited to reveal which is one is better than the other.
References
1. Munro, P.: Shared network resources and shared task properties. In: Proceedings ofthe Eighteenth Annual Conference of the Cognitive Science Society (1996)
2. Munro, P., Bao, J.: A connectionist implementation of identical elements. In: Pro-ceedings of the Twenty Seventh Ann. Conf. Cognitive Science Society Proceedings.Lawerence Erlbaum: Mahwah NJ (2005)
3. Munro, P.W.: Learning structurally analogous tasks. In: Kurkova, V., Neruda, R.,Koutnık, J. (eds.) Artificial Neural Networks - ICANN 2008, 18th InternationalConference. Lecture Notes in Computer Science, vol. 5164, pp. 406–412. Springer:Berlin/Heidelberg (2008)
4. Munro, P., Peng, Y.: Analogical learning and inference in overlapping networks. In:Kokinov, B., Holyoak, K., Gentner, D. (eds.) New Frontiers in Analogy Research.New Bulgarian University Press (2009)
5. Peng, Y., Munro, P., Mao, M.: Ontology mapping neural network: An approach tolearning and inferring correspondences among ontologies. In: Proceedings of the 9thInternational Semantic Web Conference (2010)
240
Towards Tailored Domain Ontologies
Cheikh Niang1,2, Beatrice Bouchou
1, and Moussa Lo2
1 Universite Francois Rabelais Tours - Laboratoire d’Informatique2 Universite Gaston Berger de Saint-Louis - LANI
Introduction. The goal of domain ontology is to provide a common concep-tual vocabulary to members of a virtual community of users who need to sharetheir information in a particular domain (such as medical, tourism, banking,agricultural). The identification and definition of concepts that describe the do-main knowledge requires a certain consensus. Generally, each member or sub-community holds some knowledge, he has its own view on the domain, and hedescribes it with his own vocabulary. Thus, to reach a consensus allowing toreflect a common view of the domain can be a difficult task and even moreharder if members are geographically dispersed. One way very widely used is tostart from pre-existent elements in the domain: text corpus, taxonomies, ontol-ogy fragments, and to exploit them as a basis for gradually defining the domainontology [2][7].
In this short paper, we present an approach using Ontology Matching tech-niques [1][5][6][3] for building a tailored domain ontology, starting from a generaldomain taxonomy and several pieces of knowledge given by different partners.
Our strategy is to design a mediator, firstly to reach an agreement witheach partner on their knowledge fragments that will be part to the shared do-main ontology, and secondly to conciliate these various fragments by linking andstructuring the concepts that compose them. As a mediator ontology, in ourcase study we use a public taxonomy that exists for describing subject fieldsin agriculture, forestry, fisheries, food and related domains (e.g. environment),called AGROVOC3. The resulting domain ontology combines the following twofeatures: (i) it is the portion of the general taxonomy that is relevant to theconsidered application domain as seen by each partner, (ii) it is completed andtailored by relations and properties coming from partner’s data. Fig. 1 showsan example of a domain ontology DO built starting from two local ontologiesLO1 and LO2. DO’s concepts prefixed with ag are from AGROVOC. One cansee that in DO Plan products and Varieties are related and also that they arerelated to attributes price and surface, which is not the case in AGROVOC.
Reaching an agreement with a partner. This is the first step of our generalapproach. Each partner’s fragment knowledge is represented by a Local On-tology, denoted by LO. The agreement between the mediator and the partneris concluded based on a matching between LO and the mediator ontology MO.It is consented by the partner that each concept of LO which can be associatedwith a concept of MO, called its anchor, will be a concept of the tailored domain
3 http ://www4.fao.org/agrovoc/4 This work is supported by ANR-08-DEFIS-04
241
sd:Ricee
sd:Tomatosd:Varieties
ag:Plan_products ag:Varieties
LO1
DO
ag:Tomato ag:Vegetables
ag:Onion
surface literalag:Rice
ag:Cereals
ag:Sorghum
ad:Sorghume
ad:Onionad:Varieties
price literal
price literal
surface literal
price literal
price literal
surface literal
surface literal
LO2
Fig. 1. Domain Ontology built starting from LO1 and LO2.
ontology DO. This agreement is also an ontology composed by the anchored con-cepts of LO with their anchor, as well as the local relationships between them.
Conciliation. Once the mediator has found an agreement with each partner onthe concepts which must be part to the domain ontology, it applies a conciliationphase at the end of which the domain ontology is built. This is an incrementalyphase, the local ontologies are conciliated by integrating their agreement into thedomain ontology DO, one after another. To achieve efficiently this phase, (i) themediator ontology is partitioned into blocks, according to Falcon-AO method [4]and (ii) conflict resolution strategies are applied. Each block is a sub-ontologyof MO containing semantically close concepts. Our algorithm relies on this clas-siffication in order to find links that exist between the concepts already presentin the domain ontology and those of the new local ontology to conciliate.
References
1. Zharko Aleksovski, Warner ten Kate, and Frank van Harmelen. Exploiting the struc-ture of background knowledge used in ontology matching. In Ontology Matching,2006.
2. Loris Bozzato, Mauro Ferrari, and Alberto Trombetta. Building a domain ontologyfrom glossaries: A general methodology. In SWAP, 2008.
3. Jerome Euzenat and Pavel Shvaiko. Ontology matching. Springer, 2007.4. Wei Hu, Yuanyuan Zhao, and Yuzhong Qu. Partition-based block matching of large
class hierarchies. In ASWC, pages 72–83, 2006.5. Yannis Kalfoglou and Marco Schorlemmer. Ontology mapping: The state of the art.
The Knowledge Engineering Review, 18:2003, 2003.6. Pavel Shvaiko and Jerome Euzenat. Ten challenges for ontology matching. In
Proceedings of The 7th International Conference on Ontologies, DataBases, and
Applications of Semantics (ODBASE), 2008.7. Jie Yang, Lei Wang, Song Zhang, Xin Sui, Ning Zhang 0003, and Zhuoqun Xu.
Building domain ontology based on web data and generic ontology. In Web Intelli-
gence, pages 686–689, 2004.
242
Crowd sourcing through social gaming for community driven ontology engineering, results and observations
Alloy Martin Chua, Roland Christian Chua, Arthur Vincent Dychiching, Tinmon Ang, Jose Lloyd Espiritu, Nathalie Rose Lim, Danny C. Cheng
College of Computer Studies, De La Salle University Manila Philippines {alchua01, arvolution7000, avd_arthur, angtinmon}@yahoo.com,
[email protected], [email protected], [email protected]
Abstract. In developing ontology, expert driven approaches lack the scalability to accommodate the vast amount of data on the web. As such, the community is being tapped to build ontologies to cope with highly dynamic data sources. Common problems (like difficulty of the task, quality of output, and incentives needed to motivate the community), as discussed by other authors, are considered. In this paper, we discuss observations on our approach to improve the quality and sustain community ontology refinement though the use of social gaming and interaction. Current observations show that profile and knowledge of the concept in question, understanding and expressivity of the relationships play a key role in the quality of the result.
Keywords: Matcher selection, self-organization, community driven ontology engineering, social gaming, incentives.
1 Introduction
Given the disadvantages of expert driven ontology engineering [1], communities of stakeholders would have to be involved in the engineering process to allow the capture of emergent data and concepts and keep pace with ontology evolution. Social Networks are suitable for this process as the members all share common background knowledge, goals, and interests. Researches by [2][3][4][6] have explored community-driven approaches or games-with-a-purpose to develop the ontology. However, these works do not consider the familiarity and perception of the user and their ability to provide quality feedback (with regards to the concept in question). Also, the motivation or incentive for the user to continuously provide input to the system and its sustainability in terms of application propagation and social influence is not fully considered. Our methodology is to present the engineering task as a Facebook game to verify a lightweight ontology extracted from delicious and consider common background knowledge or familiarity to concept. Different aspects of social influence (both direct and indirect) [5] are used to allow for sustainability and scalability of the system. In terms of sustainability, we refer to direct influence such as friend requests to participate, while for indirect influence, we look at general awareness of peer activities via public postings. For scalability, the community is
243
allowed to engineer the ontology and version the ontology on a community basis (small world graphs) and it uses direct and indirect peer influence to allow self monitoring and propagation of the application. For both cases, activities that promote social influence are tightly integrated to incentive schemes to motivate the community to perform those actions.
2 Results and Observations
During our two-month testing with 110 users subdivided into 4 community groupings, we encountered several issues and discovered some interesting results. It was seen that groups who are familiar with the area (topic) were able to participate more. These were also the users who, generally, return more than once; as opposed to the community that is from a different background. This result serves as a promising response to the assertion that selection of participants is important and knowledge of the tag in question by the community is necessary for them to provide feedback. Also, users are more capable of identifying erroneous relationships as opposed to validating
nitial discovered ontology. Perception and interpretation affect how concepts will be organized. Limiting our current discussion to simply identifying hierarchical
- , it can be seen that connections validated and invalidated differ between groups. For the non-computer science community, some users perceived video as a - learning mechanism (when they meant to express
). This was due to the lack of expressivity of the relationships in the game. It should be noted that no - relationship between
was produced from inputs in the computer science user community. It was observed perspectives and interpretation play key roles and user profile and social affinity can serve as some of the bases to limit and select users for participation. Finally, the ontology is able to stabilize and converge; however, it is still not known whether it is the stability of the domain that allowed the ontology to converge or the process itself.
References
1. Dieter Fensel. Ontology Management Semantic Web, Semantic Web Services, and Business Applications. [ed.] Pieter De Leenheer, Aldo de Moor, York Sure Martin Hepp. New York : Springer Science+Business Media, LLC, 233 Spring Street, New York, NY, 2008.
2. Siorpaes, Katharina, Prantner, Kathrin and Bachlechner, Daniel. Class Hierarchy for the e-tourism Ontology Version 8. e-Tourism. [Online] November 2004. [Cited: January 6, 2009.] http://e-tourism.deri.at/ont/.
3. Cardoso, Jorge and Sheth, Amit P. Semantic Web Services, Processes and Applications. United States of America : Springer Science+Business Media, LLC, 2006.
4. Steve LEUNG, Fuhua LIN, Dunwei WEN. Towards a Data-driven Ontology Engineering. Taipei : The 16th International Conference on Computers in Education, 2008.
5. Eytan Bakshy, Brian Karrer,Lada A. Adamic. Social Influence and the Diffusion of User-Created Content. California : ACM, 2009. pp. 325-334. ISBN:978-1-60558-458-4.
6. Hepp, Katharina Siorpaes and Martin. OntoGame: Weaving the Semantic Web by Online Gaming. Tenerife : Springer, 2008.
244